Applied Computational Fluid Dynamics (Mechanical Engineering Series) (Dekker Mechanical Engineering)

MwBl€D COMPlJlRllONRl FllJlD DYNAMICS

MECHANICAL ENGINEERING A Series of Textbooks and Reference Books

Editor

L. L. Faulkner Columbus Division, Battelle Memorial Institute and Department of Mechanical Engineering The Ohio State University Columbus, Ohio

1. Spring Designer's Handbook, Harold Carlson 2. Computer-Aided Graphics and Design, Daniel L. Ryan 3 . Lubrication Fundamentals, J. George Wills 4. Solar Engineering for Domestic Buildings, William A. Himmelman 5. Applied Engineering Mechanics: Statics and Dynamics, G . Boothroyd and C. Poli 6. Centrifugal Pump Clinic, lgor J. Karassik 7. Computer-Aided Kinetics for Machine Design, Daniel L. Ryan 8. Plastics Products Design Handbook, Part A: Materials and Components; Part B: Processes and Design for Processes, edited by Edward Miller 9. Turbomachinery: Basic Theory and Applications, Earl Logan, Jr. 10. Vibrations o f Shells and Plates, Werner Soedel 11. Flat and Corrugated Diaphragm Design Handbook, Mario Di Giovanni 12. Practical Stress Analysis in Engineering Design, Alexander Blake 13. An Introduction to the Design and Behavior of Bolted Joints, John H. Bic kf ord 14. Optimal Engineering Design: Princbles and Applications, Jarnes N. Siddall 15. Spring Manufacturing Handbook, Harold Carlson 16. Industrial Noise Control: Fundamentals and Applications, edited by Lewis H. Bell 17. Gears and Their Vibration: A Basic Approach to Understanding Gear Noise, J. Derek Smith 18. Chains for POwer Transmission and Material Handling: Design and Applica tions Handbook , American Chain Association 19. Corrosion and Corrosion Protection Handbook, edited by Philip A. Schweitzer 20. Gear Drive Systems: Design and Application, Peter Lynwander 21. Controlling In-Plant Airborne Contaminants: Systems Design and Calculations, John D. Constance 22. C A D K A M Systems Planning and Implementation, Charles S. Knox 23. Probabilistic Engineering Design: Principles and Applications, James N. Siddall

24. Traction Drives: Selection and Application, Frederick W. Heilich Ill and Eugene E. Shube 25. Finite Element Methods: An Introduction, Ronald L. Huston and Chris E. Passerello 26. Mechanical Fastening of Plastics: An Engineering Handbook, Brayton Lincoln, Kenneth J. Gomes, and James F. Braden 27. Lubrication in Practice: Second Edition, edited by W. S . Robertson 28. Principles of Automated Drafting, Daniel L. Ryan 29. Practical Seal Design, edited by Leonard J. Martini 30. Engineering Documentation for CAD/CAM Applications, Charles S . Knox 3 1 . Design Dimensioning with Computer Graphics Applications, Jerome C. Lange 32. Mechanism Analysis: Simplified Graphical and Analytical Techniques, Lyndon 0 . Barton 33. CAD/CAM Systems: Justification, Implementation, Productivity Measurement, Edward J. Preston, George W. Crawford, and Mark €. Coticchia 34. Steam Plant Calculations Manual, V. Ganapathy 35. Design Assurance for Engineers and Managers, John A. Burgess 36. Heat Transfer Fluids and Systems for Process and Energy Applications, Jasbir Singh 37. Potential Flows: Computer Graphic Solutions, Robert H. Kirchhoff 30. Computer-Aided Graphics and Design: Second Edition, Daniel L. Ryan 39. Electronically Controlled Proportional Valves: Selection and Application, Michael J. Tonyan, edited by Tobi Goldoftas 40. Pressure Gauge Handbook, AMETEK, U.S. Gauge Division, edited by Philip W. Harland 4 1. Fabric Filtration for Combustion Sources: Fundamentals and Basic Technology, R. P. Donovan 42. Design of Mechanical Joints, Alexander Blake 43. CAD/CAM Dictionary, Edward J. Preston, George W. Crawford, and Mark E. Coticchia 44. Machinery Adhesives for Locking, Retaining, and Sealing, Girard S . Haviland 45. Couplings and Joints: Design, Selection, and Application, Jon R. Mancuso 46. Shaft Alignment Handbook, John Piotrowski 47. BASIC Programs for Steam Plant Engineers: Boilers, Combustion, Fluid Flow, and Heat Transfer, V. Ganapathy 40. Solving Mechanical Design Problems with Computer Graphics, Jerome C. Lange 49. Plastics Gearing: Selection and Application, Clif ford E. Adams 50. Clutches and Brakes: Design and Selection, William C . Orthwein 5 1 Transducersin Mechanical and Electronic Design, Harry L. Trietley 52. Metallurgical Applications of Shock-Wave and High-Strain-Rate Phenomena, edited by Lawrence E. Murr, Karl P. Staudhammer, and Marc A. Meyers 53. Magnesium Products Design, Robert S. Busk 54. How to Integrate CAD/CAM Systems: Management and Technology, William D. Engelke

5 5 . Cam Design and Manufacture: Second Edition; with cam design software for the IBM PC and compatibles, disk included, Preben W. Jensen 56. Solid-state AC Motor Controls: Selection and Application, Sylvester Campbell 57. Fundamentals of Robotics, David D. Ardayfio 58. Belt Selection and Application for Engineers, edited by Wallace D. Erickson 59. Developing Three-Dimensional CAD Software with the IBM PC, C . Stan W ei 60. Organizing Data for CIM Applications, Charles S . Knox, with contributions by Thomas C. Boos, Ross S. Culverhouse, and Paul F. Muchnicki 61. Computer-Aided Simulation in Railway Dynamics, by Rao V. Dukkipati and Joseph R. Amyot 62. Fiber-Reinforced Composites: Materials, Manufacturing, and Design, P. K . Mallick 63. Photoelectric Sensors and Controls Selection and Application, Scott M. Juds 64. finite Element Analysis with Personal Computers, Edward R. Champion, Jr . , and J. Michael Ensminger 6 5. Ultrasonics: Fundamentals, Technology, Applications: Second Edition, Revised and Expanded, Dale Ensminger 66. Applied Finite Element Modeling: Practical Problem Solving for Engineers, Jeffrey M. Steele 67. Measurement and Instrumentation in Engineering: Principles and Basic Laboratory Experiments, Francis S. Tse and lvan E. Morse 68. Centrifugal Pump Clinic: Second Edition, Revised and Expanded, lgor J. Karassik 6 9. Practical Stress Analysis in Engineering Design: Second Edition, Revised and Expanded, Alexander Blake 70. An Introduction to the Design and Behavior of Bolted Joints: Second Edition, Revised and Expanded, John H. Bickford 71. High Vacuum Technology: A Practical Guide, Marsbed H. Hablanian 72. Pressure Sensors: Selection and Application, Duane Tandeske 73. Zinc Handbook: Properties, Processing, and Use in Design, Frank Porter 74. Thermal Fatigue of Metals, Andrzej Weronski and Tadeusz Hejwowski 75. Classical and Modern Mechanisms for Engineers and Inventors, Preben W. Jensen 76. Handbook of Electronic Package Design, edited by Michael Pecht 77. Shock- Wave and High-Strain-Rate Phenomena in Materials, edited by Marc A. Meyers, Lawrence E. Murr, and Karl P. Staudhammer 78, Industrial Refrigeration: Principles, Design and Applications, P. C . Koelet 79. Applied Combustion, Eugene L. Keating 80. Engine Oils and Automotive Lubrication, edited by Wilfried J. Bartz 8 1. Mechanism Analysis: Simplified and Graphical Techniques, Second Edition, Revised and Expanded, Lyndon 0 . Barton 82, Fundamental Fluid Mechanics for the Practicing Engineer, James W. Murdock 8 3. Fiber-Reinforced Composites: Materials, Manufacturing, and Design, Second Edition, Revised and Expanded, P. K. Mallick

84. Numerical Methods for Engineering Applications, Edward R. Champion, Jr. 8 5. Turbomachinery: Basic Theory and Applications, Second Edition, Revised and Expanded, Earl Logan, Jr. 86. Vibrations of Shells and Plates: Second Edition, Revised and Expanded, Werner Soedel 87. Steam Plant Calculations Manual: Second Edition, Revised and Ex panded, V. Ganapathy 88. Industrial Noise Control: Fundamentals and Applications, Second Edition, Revised and Expanded, Lewis H. Bell and Douglas H. Bell 89. Finite Elements: Their Design and Performance, Richard H. MacNeal 90. Mechanical Properties of Polymers and Composites: Second Edition, Revised and Expanded, Lawrence E. Nielsen and Robert F. Landel 9 1. Mechanical Wear Prediction and Prevention, Raymond G . Bayer 92. Mechanical Power Transmission Components, edited by David W. South and Jon R. Mancuso 93. Handbook of Turbomachinery,edited by Earl Logan, Jr. 94. Engineering Documentation Control Practices and Procedures, Ray E. Monahan 95. Refractory Linings Thermomechanical Design and Applications, Charles A . Schacht 9 6. Geometric Dimensioning and Tolerancing: Applications and Techniques for Use in Design, Manufacturing, and Inspection, James D. Meadows 97. An Introduction to the Design and Behavior of Bolted Joints: Third Edition, Revised and Expanded, John H. Bickford 98. Shaft Alignment Handbook: Second Edition, Revised and Expanded, John Pi0t ro w s ki 99. Computer-Aided Design of Polymer-Matrix Composite Structures, edited by Suong Van Hoa 100. Friction Science and Technology, Peter J. Blau 10 1. Introduction to Plastics and Composites: Mechanical Properties and Engineering Applications, Edward Miller 102. Practical Fracture Mechanics in Design, Alexander Blake 103. Pump Characteristics and Applications, Michael W. Volk 104. Optical Principles and Technology for Engineers, James E. Stewart 105. Optimizing the Shape of Mechanical Elements and Structures, A. A. Seireg and Jorge Rodriguez 106. Kinematics and Dynamics of Machinery, Vladimir Stejskal and Michael ValdSek 107. Shaft Seals for Dynamic Applications, Les Horve 108. Reliability-Based Mechanical Design, edited b y Thomas A . Cruse 109. Mechanical Fastening, Joining, and Assembly, James A . Speck 110. Turbomachinery Fluid Dynamics and Heat Transfer, edited by Chunill Hah 11 1 . High- Vacuum Technology: A Practical Guide, Second Edition, Revised and Expanded, Marsbed H. Hablanian 1 12. Geometric Dimensioning and Tolerancing: Workbook and Ans werbook, James D. Meadows 1 13. Handbook of Materials Selection for Engineering Applications, edited by G. 1.Murray

1 14. Handbook of Thermoplastic Piping System Design, Thonias Sixsmith and Reinhard Hanselka 1 15. Practical Guide to Finite Elements: A Solid Mechanics Approach, Steven M. Lepi 1 16. Applied Computational Fluid Dynamics, edited by Vijay K. Garg

Additional Volumes in Preparation Fluid Sealing Technology, Heinz Konrad Muller and Bernard Nau Friction and Lubrication in Mechanical Design, A. A. Seireg Machining of Ceramics and Composites, edited by Said Jahanmir and M. Ramulu Heat Exchange Design Handbook, T . Kuppan Couplings and Joints, Jon R. Mancuso

Mechanical Engineering Software Spring Design with an I6M PC, AI Dietrich Mechanical Design Failure Analysis: With Failure Analysis System Software for the l6M PC, David G. Ullman

RPPll€D COMPUTfiTlONRl FLUID DYNRMICS €DlT€D BY VIJRY K. GRRG

AYT Corporation/NASALewis Research Center Cleveland, Ohio

MARCEL

MARCELDEKKER, INC. D E K K E R

NEWYORK BASEL HONGKONG

Library of Congress Cataloging-in-Publication Data Applied computational fluid dynamics / edited by Vijay K. Garg. p. cm. -- (Mechanical engineering ; 116) Includes index. ISBN 0-8247-0 165-8 (alk. paper) 1. Fluid dynamics. 2. Heat--Transmission. 3. Turbulence. I . Garg, Vijay K. 11. Series: Mechanical engineering (Marcel Dekker, Inc.) ; 1 16. TA357.A657 1998 620.It064--dc21

98- 16640 CIP

This book is printed on acid-free paper.

Headquarters Marcel Dekker, Inc. 270 Madison Avenue, New York, NY 10016 tel: 2 12-696-9000: fax: 2 12-685-4540 Eastern Hemisphere Distribution Marcel Dekker AG Hutgasse 4,Postfach 8 12, CH-4001 Basel, Switzerland tel: 44-6 1-261-8482: f a : 44-6 1-261-8896 World Wide Web http://www.dekker.com 'The publisher ofTers discounts on this book when ordered in bulk quantities. For more information, write to Special SaledProfessional Marketing at the headquarters address above.

Copyright 0 1998 by Marcel Dekker, Inc. All Rights Reserved. Neither this book nor any part may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, microfilming, and recording, or by any information storage and retrieval system, without permission in writing from the publisher. Current printing (last digit): 1 0 9 8 7 6 5 4 3 2 1

PRINTED IN THE UNITED STATES OF AMERICA

Preface

In recent years, growth in computational capability has led to a phenomenal increase in the use of computational methods for engineering design, especially for the design of fluid flow and thermal systems. For such systems, the governing equations are too complex to be solved analytically except in some trivial cases. Thus the use of numerical techniques for problems of practical interest in thermal systems has mushroomed in the last two decades. With ever-increasing costs of experimentation, the need for numerical simulation of the relevant processes is even greater. Of course, experimental data are indispensable for checking the accuracy and validating the numerical model. Computational fluid dynamics (CFD) has become a major tool for designers of fluid flow and thermal systems. Progress in this area has been rapid, and the use of three-dimensional methods is increasingly applied to the design process. The availability of faster and larger computers has enabled us to solve complex fluid flow and thermal problems that arise in a wide variety of applications. One often finds a lack of valuable applications when teaching CFD to students from a variety of disciplines. Most examples in textbooks on CFD either reach and appeal to few students outside the specific discipline for which the textbook was written or contain rather simple applications. Part of the philosophy behind the preparation of this book was to expose the reader to the latest computational techniques used for the solution of reallife problems in the field of fluid flow and heat transfer. It is our intention that this book, although self-contained, will not be used in isolation. For iii

Preface

the prospective student of CFD, there is no substitute for extensive reading and knowledge of the latest developments in order to obtain a true understanding of the subject. The principal objective in this book has been to bring together a collection of applications of C F D that will be beneficial to workers in diverse fields, such as environmental sciences, energy systems, mechanical engineering, chemical engineering, and aerospace. For each application, there is a contribution from experts who have worked in the particular field for 15-20 years and are at the forefront of the techniques used. This book explores real-life applications of computational fluid dynamics and heat transfer in various fields such as aeronautics, materials processing and manufacturing, electronic cooling, and environmental control. For example, it enhances the reader’s knowledge of how to utilize the modern tools of C F D to increase the efficiency, reduce the fuel consumption, and increase the affordability of aircraft engines ; manufacture and process better materials; improve the quality of the air we breathe; and cool computers better so that they can run faster and cheaper, and thus help us solve more complex problems in a shorter time. The reader is exposed to the state of the art in these applications of C F D techniques. In order to provide the mathematical background for this subject, Chapter 1 begins with a discussion of the nature of governing equations along with the relevant boundary conditions that arise in fluid flow and heat transfer problems. This chapter also describes the governing equations in general curvilinear coordinates, and discusses the mathematical properties of these equations. Chapter 2 describes the three basic techniques-the finite difference method, the finite volume method, and the finite element method-along with considerations of accuracy, stability, and convergence. This is followed by a state-of-the-art review of turbulence modeling, both dynamic and thermal, in Chapter 3. Following the development of basic concepts, this chapter describes the various eddy viscosity and stress transport models for dynamic turbulence and then discusses the various thermal models. This background is supplemented with grid generation techniques in Chapter 4. In fact, C F D has provided much of the impetus for the development of these techniques, which are equally applicable to all physical problems involving field solutions. This chapter describes grid generation via the solution of partial differential equations as well as algebraic, adaptive, and unstructured grid generation techniques. It also provides some details for the various codes currently available for grid generation. The various applications are then discussed in individual chapters as described below. Chapter 5 discusses inlet, duct, and nozzle flows. After describing the CFD solution process for such flows, it provides details on the application

Preface

of the process to two examples. The chapter closes with the current status and future directions in terms of modeling, numerical, and procedural issues. Chapter 6 looks at the impact of unsteadiness in turbine flows with a view to distinguish flow parameters that can be modeled with existing steady CFD codes from those that require unsteady codes. This is an important issue since unsteady three-dimensional computation of heat transfer in complex geometries such as a turbomachine is prohibitively expensive, even with present-day computers. Chapter 7 discusses the numerical modeling of heat transfer and fluid flow processes in the thermal processing of materials. It brings out the importance of material properties in an accurate modeling of the process. This chapter also discusses the numerical results and computational problems that arise in various processes, such as plastic extrusion, optical fiber drawing, casting, and heat treatment. Application of CFD techniques in electronic cooling is discussed in Chapter 8. Cost, size, and weight are often the primary constraints on the thermal and physical design of electronic products used in computers; military/aerospace, industrial, and consumer products ; business/retail and automotive markets; instrumentation; and telecommunications. Passive thermal control techniques are discussed together with examples. Chapter 8 closes with some thoughts on future directions on the use of C F D for electronic cooling. Chapter 9 discusses the application of CFD to control air quality. Problems such as acid deposition, smog, global climate warming, and stratospheric ozone depletion pose significant threats to both human health and welfare and related ecological damage. The role of air quality models is to identify effective solutions to the complex environmental problems. This chapter describes several advection schemes used by air quality models, and evaluates them for various test cases against the different performance measures. The chapter authors and I will consider our efforts successful if readers are able to apply CFD techniques to the resolution of problems in their areas of concern. I wish to thank Dr. Raymond E. Gaugler of NASA Lewis Research Center for his support. Thanks are also due to the staff at Marcel Dekker, Inc., for their help and competent handling of this project. My family has been very understanding and supportive during the writing activity, and I am indebted to them. Vijap K . Garg

This page intentionally left blank

Contents

Preface Contributors

Governing Equations Vijap K . Garg

...

111

is 1

Numerical Techniques Vijap K . Garg

Turbulence Modeling John R . Schwab

Grid Generation Vijay K . Garg and Philip C . E. Jorgenson

117

Inlet, Duct, and Nozzle Flows Charles E. Towne

179

Turbine Flows : The Impact of Unsteadiness O m P. Sharma, Daniel J . Dorney, Seyf Tanrikut, and Ron-Ho N i

22 1 vi i

Contents

viii

Numerical Modeling of Materials Processing and Manufacturing Systems Yogesh Jaluria

Passive Thermal Control of Electronic Equipment Yoyendra Joshi

Computational Fluid Dynamic Techniques in Air Quality Modeling Mehmet T . Odman and Armistead G. Russell

257 30 1

335

Appendix :Governing Equations in Various Systems

403

Index

41 7

Contributors

Daniel J. Dorney, Ph.D.* Project Engineer, Pratt & Whitney Aircraft, East Hartford, Connecticut Vijay K. Garg, Ph.D, Senior Research Engineer, AYT Corporation/NASA Lewis Research Center, Cleveland, Ohio Yogesh Jaluria, Ph.D. Professor, Department of Mechanical Engineering, Rutgers University, New Brunswick, New Jersey Philip C. E. Jorgenson, Ph.D. Aerospace Engineer, Engine Systems Technology Branch, NASA Lewis Research Center, Cleveland, Ohio Yogendra Joshi, Ph.D. Associate Professor, Department of Mechanical Engineering, University of Maryland, College Park, Maryland

* Current

affiliation : Assistant Professor, Department of Mechanical Engineering, G M I Engineering & Management Institute, Flint, Michigan. ix

Contributors

Ron-Ho Ni, Ph.D. Senior Feiiow, Pratt & Whitney Aircraft, East Hartford, Connecticut Mehmet T. Odman, Ph.D.* Research Scientist, MCNC--Environmental Programs, Research Triangle Park, North Carolina Armistead G . Russel!, Ph.D. Georgia Power Professor: Department of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, Georgia John R. Schwab Fluid and Thermal Analysis Consultant, North Olmsted, Ohio Om P. Sharma, Ph.D. Senior Fellow, Pratt & Whitney Aircraft, East Hart ford. Connecticut Seyf Tanrikut, Ph.D. Senior Design Engineer, Prat,! & Whitney Aircraft, East Hartford. Connecticut Charles E. Tnwne, Ph.D. Aerospace Engineer, NASA Lewis Research Center, Cleveland, Ohio

_ ~ _

____

* Current czfj’iliution:Senior Research Engineer, Department of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, Georgia.

1 Governing Equations .

Vijay K Garg A Y T CorporationjN A S A Lewis Research Center. Cleveland. Ohio

1.1 1.2 1.3 1.4 1.5 1.6 1.7

Continuity Equation . . . . . Momentum Equation . . . . . Energy Equation . . . . . . . Boundary Conditions . . . . . Equation of State . . . . . . Vector Form of Equations . . . Nondimensional Form of Equations

. . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . 1.8 Boundary Layer Equations . . . . . .

. . . . . . . . . . . .

. . . . . .

1.8.1 Boundary Layer Equations for Incompressible. Two-Dimensional Flow . . . . . . . . . 1.8.2 Boundary Layer Equations for Compressible Flow 1.9 Euler Equations . . . . . . . . . . . . . . . 1.9.1 Continuity Equation . . . . . . . . . . 1.9.2 Inviscid Momentum Equations . . . . . . . 1.9.3 Inviscid Energy Equations . . . . . . . . 1.9.4 Simplified Form of Euler Equations . . . . . 1.10 Parabolized Navier-Stokes Equations . . . . . . 1.11 Governing Equations in Generalized Coordinates . .

. . . . . .

. . . . .

. . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . .

2 3 6 7 8 9 10 11

12 13 16 16 18 20 20 21 25 1

Garg

2 1.12 Mathematical Properties of the Governing Equations . . . . . . .

Nomenclature . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . .

31 32

The equations governing the flow of a fluid and the associated heat transfer are based on the conservation principles for mass, momentum, and energy. These equations are first presented for a compressible, viscous, Newtonian fluid, and then particularized for simpler cases. It is assumed that the reader has some background in this field. Thus, a complete derivation of the governing equations is not included. The reader is referred to Schlichting (1979) for the derivation. For the general case of three-dimensional motion, the flow field is specified by the velocity vector V

+ vj + wk

where U, L', U' are the three orthogonal components, by the pressure p , density p, and temperature T , all conceived as functions of the space coordinates and time t . For the determination of these six quantities, there exist six equations: the continuity equation (conservation of mass), the three equations of motion (conservation of momentum), the energy equation (conservation of energy), and the thermodynamic equation of state p = p ( p , T).

1.1

CONTINUITY EQUATION

The continuity equation implies a balance between the masses entering and leaving a control volume per unit time and the change in density within it. For the unsteady flow of a compressible fluid, the conservation of mass applied to a fluid passing through an infinitesimal, fixed control volume yields the following equation of continuity: 8P

DP + p(V + v * ( p y )= Dt

v) = 0

The first term in this equation represents the rate of increase of density in the control volume and the second term represents the rate of mass flux passing out of the control surface (which surrounds the control volume) per

Governing Equations

unit volume. The symbol Dp/Dt denotes the substantive derivative, which consists of the local contribution (in unsteady flow) dp/& and the convective contribution (due to translation) V *(Vp). Equation (1.1) is based on the Eulerian approach, in which changes to the fluid are recorded as the fluid passes through a fixed control volume. In the alternative Lagrangian approach, changes to the properties of a fluid element are recorded by an observer moving with the fluid element. In general, the Eulerian approach is used for fluid flows. A flow in which the density of each fluid element remains constant is called incompressible. This implies that Dp/Dt = 0, which reduces Eq. (1.1) to

for a steady or unsteady incompressible flow. The assumption of incompressibility is a good approximation for air flows when the Mach number is less than 0.3. In indicial notation, Eq. (1.1) can be written as

1.2

MOMENTUM EQUATION

The momentum equation is derived from Newton’s second law of motion. Further, in fluid motion it is necessary to consider two types of forces separately: (1) forces acting throughout the mass of the fluid element (e.g., gravitational forces), known as body forces, and (2) forces acting on the boundary (pressure and friction), known as surface forces. If Fi denotes the body force per unit volume, and Pi denotes the surface force per unit volume, both in direction, i, the momentum equation in indicial notation can be written as p

DV = Fi + P i , Dt

i = 1,2,3

where D K / D t denotes the substantive acceleration of the fluid element. The most common body force is the gravitational force. In this case F , = p g i , where gi is the acceleration due to gravity. While the body forces are regarded as given external forces, the surface forces depend on the rate at

Garg

which the fluid is strained by the velocity field in it. In fact, the surface force Piis related to the stress (force per unit area) by

where oij is the stress on U plune nornzul to the i axis in the direction of t h e j axis. The stress tensor is symmetric, i.e., aij = a j , , except if body moments are present as in the case of a magnetic body in a magnetic field. The momentum equation given is quite general and is applicable to both continuum and noncontinuum flows. It loses its generality, however, when approximate expressions are inserted for the stress tensor. Here we restrict ourselves to isotropic, Newtonian fluids for which the stress at a point is linearly related to the rate of strain (deformation) of the fluid. All gases which can be treated as a continuum, and most liquids of interest, in particular water, belong to this class. For such fluids, the so-called constitutive equations (relation between stress and rutt' c f s t r a i n ) are (Round and Garg, 1986)

where the Kronecker delta dij = 0 for i # j and dij = 1 for i = j , p is the coefficient of viscosity (dynamic viscosity), and p' is the second coefficient of viscosity. The two coefficients of viscosity are related to the coefficient of bulk viscosity ( K ) by the expression 2 3

u=-p+p'

( 1.7)

Except in the study of the structure of shock waves, and in the absorption and attenuation of acoustic waves, it is generally believed that the coefficient of bulk viscosity is negligible. With K = 0, the second coefficient of viscosity becomes p , = - - 2p

and the stress tensor may be written as

The stress tensor is also written as (1.10)

Governing Equations

where zij represents the viscous (shear) stress tensor given by the bracketed term on the right side of Eqs. (1.9). It is easily seen from Eqs. (1.6) that, for an incompressible fluid, the mean normal stress is equal to the negative pressure; i.e., 1

3 Oii

due to the continuity equation (1.2). Substituting (1.5) and (1.9) into (1.4) yields

(i,j, k

1, 2, 3)

(1.11)

These well-known differential equations form the basis of the whole science of fluid mechanics, and are known as the Nacier-Stokes equations. In this form, they are applicable to a viscous, compressible, isotropic, Newtonian fluid with variable properties. In general, the viscosity p may be regarded as dependent on the space coordinates, since p varies considerably with temperature (though little with pressure). In such a case, the temperature dependence of viscosity p ( T )must be obtained from experiments. For an incompressible fluid, the last term in Eq. (1.1 1) vanishes identically due to the continuity equation (1.2). Further, since temperature variations are, generally speaking, small in this case, the viscosity may be assumed to be constant.* With this assumption, Eq. (1.11) simplifies to p ( 2 +

y % ) = F i - - aP axj

axi

dxjdxi

(i, j , k = 1, 2, 3)

(1.12)

In vector notation, Eq. (1.12) can be written as DV Dt

p - = F - Vp -I- pV2 V

(1.13)

For an intliscid fluid, this reduces to (1.14)

* This condition is more nearly satisfied in gases than in liquids.

Garg

which is the well-known Euler equation. It is valid for compressible and incompressible inviscid flows. 1.3 ENERGY EQUATION

The energy equation is derived from the first law of thermodynamics, and in the absence of chemical reaction and radiation it can be written in indicial notation as

vg)+

p(:+

av.

p axi = axi ( k

E)+

(1.15)

where e is the specific internal energy, k is the thermal conductivity of the fluid, and Fourier's law for heat transfer by conduction through the control surface has been assumed. Also, the total energy per unit volume, E , has been taken to be

= p(e

so that only internal energy and kinetic energy are considered significant. The dissipation function 0 represents the rate at which energy is dissipated per unit volume of fluid through the action of viscosity, and is given by

1 a& 0 = - 7..- = IJ

axj

(--)av

2 + av.av -- J

i?xi axj

(-)a&

3 dxi

(1.16)

Equation (1.15) enjoys general validity, and can be simplified for special cases. For a perfect gas, using the continuity equation (1.3) together with c,DT

c,,DT

(3 -

we can simplify Eq. (1.15) to the form DT ~t

Dp ~t

pc, - = -

i? +axi

(k ")

axi

+p0

(1.17)

where c, and c , represent the specific heats per unit mass at constant pressure and at constant volume, respectively. For an incompressible fluid, using the continuity equation (1.2) together with De = C D T simplifies Eq. (1.15) to DT 2 pc = -( k E )+ p 0 Dt 2xi ?xi

(1.18)

Governing Equations

where C is the specific heat per unit mass, and the expression for the dissipation function @ also simplifies to (1.19)

For an inuiscid fluid, p

= 0, so

Eq. (1.15) reduces to (1.20)

For an adiabatic flow of an inuiscid fluid,

p-+p-=o ~t axi

(1.21)

The left side of Eq. (1.15) can be written as

where h = e + p / p is the specific enthalpy of the fluid. Strictly speaking, the term “Navier-Stokes equations” refers to the components of the momentum equation (1.11). However, it is common practice to include the continuity and energy equations in the set of equations referred to as the Navier-Stokes equations. See Appendix A for the Navier-Stokes equations expressed in different coordinate systems. The unsteady Navier-Stokes equations govern the laminar and turbulent flows as long as effective values of the transport coeficients p and k are used. However, they are not easy to solve for turbulent flows in comparison to that for laminar flows, since the space and time scales of turbulent motion are very small. Thus, a very large number of grid points and a very small time step are required for solution. This puts the computation of many practical turbulent flows via DNS (direct numerical simulation) outside the realm of possibility for present-day computers. The main thrust in the computation of turbulent flows has been through the solution of time-averaged Navier-Stokes equations. These equations are also referred to as the Reynolds averaged equations, and are discussed in detail in Chapter 3. 1.4

BOUNDARY CONDITIONS

The solution of the foregoing equations can be determined only when the boundary and initial conditions are specified. For a viscous fluid, the condition of no slip on solid boundaries must be satisfied; i.e., on a solid, impermeable wall, the normal and tangential components of velocity must

Garg

vanish. If the energy equation is also used, temperature and/or its gradient at the boundaries should also be specified. For an iniliscid fluid, the tangential component of fluid velocity at a solid wall is not required to vanish. EQUATION OF STATE

1.5

In order to close this system of equations, it is necessary to establish relationships between the thermodynamic variables ( p , p , T , e, h ) and to relate the transport properties (p,k ) to them. For example, for a compressible flow without external heat addition, the relevant equations to be solved are Eq. (1.1) or Eq. (1.3) for the continuity equation, Eq. (1.11) for the three momentum equations, and Eq. (1.15) for the energy equation. These five equations contain seven unknowns p, p , e, T , Vl, V 2 , V 3 , assuming that the transport coefficients p, k can be related to the thermodynamic variables. Clearly, two additional equations are required to close the system. These equations are provided by the relationships among the thermodynamic variables. Such relations are known as equations of slate. From the state principle of thermodynamics, it is known that the local thermodynamic state is fixed by any two independent thermodynamic variables provided the chemical composition of the fluid does not change due to diffusion of finiterate chemical reactions. Thus, for the present example, if we choose p and T as the two independent variables, equations of state of the form P = p(p,

7-17

= e(p,

( 1.22)

are required. For a perfect gas, for example, the equation of state is p

(1.23)

pRT

where R is the gas constant. Also for a perfect gas, we have (1.24) where y is the ratio of specific heats. For air at standard conditions, R = 287 J,ikg-K and y = 1.4. For a perfect gas then, the form of Eqs. (1.22) is obvious. For fluids that cannot be considered a perfect gas, the required state relations are generally given in terms of tables, charts, or curve fits. Kinetic theory is used to relate the coeficients of viscosity and thermal conductivity to the thermodynamic variables. For example, Sutherland's law for dynamic viscosity (Schlichting, 1979) is ~ 3 ' 2

p=cl---T + C2

( 1.25)

Governing Equations

where C, and C, are constants for a given gas. For air at moderate temperatures, C , = 1.458 x 10-6 kg/(m-s-K1I2) and C2 = 110.4 K. Once p is known, the coefficient of thermal conductivity k is determined from the definition of the Prandtl number, (1.26) since the ratio cJPr is approximately constant for most gases.

1.6 VECTOR FORM OF EQUATIONS

It is often convenient to combine the governing equations into a compact vector for& before applying a numerical algorithm to them. For example, for a compressible flow without external heat addition or body forces, the governing equations in Cartesian coordinates can be written as

-dU + - + -8E + - = oaF at

(1.27)

where U, E, F,and G are vectors given by U=

ElT

PU PU2

+ P - Zxx

PUV

- Zxy

PUW

- 7x2

- UZ,, - V Z X Y - W Z X Z

+4x

PV PUV

PV2

- rxy

+ P - ryy

PVW - z y ,

- UZ,,

- vzyy - w r y ,

+ qy

1 1

(1.28)

Garg

where U , U, w represent the three components of the velocity vector V, and 4 x , yy, q, are the conductive heat fluxes in the x, y, and z directions, respectively. The first row of the vector equation (1.27) corresponds to the continuity equation (1.1) in Cartesian coordinates, the second, third, and fourth rows are the momentum equations (1.1l), while the fifth row is the energy equation (1.15).

1.7

NONDIMENSIONAL FORM OF EQUATIONS

I t is advisable to cast the governing equations in nondimensional form before carrying out a numerical solution. This enables the flow variables to be “normalized” so that their values fall between prescribed limits such as 0 and 1. Also, the characteristic parameters such as Reynolds number, Prandtl number, Mach number, etc., can be varied independently. The proper nondimensionalization is problem dependent. However, if L is the characteristic length, and other characteristic quantities are taken to be freestream values denoted by subscript CO, we may define the dimensionless variables, denoted by an asterisk, as

( 1.29)

If this nondimensionalizing procedure is applied to the compressible Navier-Stokes equations given by Eqs. (1.27) and (1.28), the following dimensionless equations are obtained :

?U* ?E* -+-+-+-=o ?t* ?.U*

ilF*

iY7

i?y*

?z*

( 1.30)

where U*, E*, P,and G* are the vectors given by Eqs. (1.28) except that each term is dimensionless, denoted by an asterisks, and the dimensionless total energy per unit volume is (1.31)

Governing Equations

The components of the shear stress tensor and the heat flux vector in dimensionless form are

z,*,=-

(

2 ~ * 2 -aw* - - - - au* 3 Re, az* ax*

qx* = -

ay*

aT*

( y - 1)ML Re, Pr

ax*

(1.32)

aT*

" = - ( y - 1 ) M i Re, 4; = -

a>.,

ay*

P* aT* ( y - 1)M: Re, Pr

Bz*

where M , and Re, are the freestream Mach number and Reynolds number, respectively, given by

and the perfect gas equations of state [Eqs. (1.22)] become

Note that the dimensionless forms of the governing equations in Eqs. (1.30) are identical (except for the asterisks) to the dimensional form in Eqs. (1.27). For convenience, the asterisks are usually dropped. 1.8 BOUNDARY LAYER EQUATIONS

Prandtl (1904) originated the concept of a boundary layer. From experimental evidence, he reasoned that when the Reynolds number is large, a

Garg

thin region (called the boundary Iaqvr) existed near a solid boundary where viscous effects were just as important as inertia effects no matter how small the viscosity of the fluid might be. Outside this boundary layer, velocity (as well as temperature) gradients were small, and since the viscosity is small or the Reynolds number is high, viscous effects are negligible. Thus, the outside mean flow pattern, being determined primarily by the boundary form. is practically that of inviscid flow past the boundary. This concept helps to reduce the governing equations considerably for the boundary layer region. Since Prandtl it has been found that a similar reduction in the governing equations is possible for flows in which a primary flow direction can be identified. Such flows include jets, wakes, mixing layers, and developing flow in pipes and other internal passages. It is thus common to refer to these reduced equations as the thin-sheur-lccjler equutions. 1.8.1

Boundary Layer Equations for Incompressible, Two-Dimensional Flow

Based on an order-of-magnitude analysis of the Navier-Stokes equations, the following boundary layer equations (in Cartesian coordinates) for an unsteady, incompressible, constant property flow over a two-dimensional body can be derived (Anderson et al., 1984): continuity : ( 1.33)

momentum : ( 1.34)

(1.35) energy :

r7T ‘:T ?T d2T -+u---+c,-=a-+--+c;t ?.U cy ?L,2

fiTui3p pc, ?.U

pc,

( 1.36)

where v is the kinematic viscosity, p / p ; a is the thermal diffusivity, k.pcs,, ; and jl is the coefficient of volumetric expansion

Governing Equations

For an ideal gas fl = 1/T, where T is the absolute temperature. The last two terms in Eq. (1.36) are retained on the assumption that the Eckert number is of order unity, which is true only for high-speed flows. The initial and boundary conditions are Initial condition : u(x, y, 0), T(x,y, 0) known everywhere No slip: u(x, 0, t ) = u(x, 0, t ) = 0

Inlet condition: u(.xo, y , t), T ( x 0 ,y, t ) known at some x 0 Patching to the outer layer: u(x, y, t ) --+ Ue(x,t ) as y + 00, T(x,y, t ) -+ T,(x, t ) as y + lx where the subscript e refers to conditions at the edge of the boundary layer. The pressure gradient term in Eqs. (1.34) and (1.36) is evaluated for the given boundary from

ap au, + U , - aue

_ -1- - -

ax -

(1.37)

This follows from the Euler equation for the inviscid outer flow. These boundary layer equations hold for flow over a curved wall as long as the boundary layer thickness is much smaller than the radius of curvature of the wall. The Reynolds averaged form of boundary layer equations for turbulent flow will be presented in Chapter 3.

1.8.2 Boundary Layer Equations for Compressible Flow

In order to reduce the Navier-Stokes equations, the order-of-magnitude analysis can also be carried out for compressible flow. Again referring the reader elsewhere for details (Cebeci and Smith, 1974), we provide the unsteady boundary layer equations for two-dimensional or axisymmetric laminar, compressible flow for the coordinate system in Fig. 1.1 : continuity : aP

a-x

- + - (r”pu) + - (r”pu) = 0

(1.38)

Garg

Figure 1.1 Coordinate system for axisymmetric boundary layer equations: ( a ) external boundary layer; (b) axisymmetric free shear layer flow; (c) confined axisymmetric flow.

momentum : (1.39) (1.40)

energy : p

c?H

i?H + pu -+ ax

i?H

pv-

3Y (1.41)

( 1.42)

Governing Equations

The enthalpy H , following boundary layer approximation, is

H=c,T+-

and n is an index equal to zero for two-dimensional flow (r" = 1) and equal to unity for axisymmetric flow (r" = r). The boundary layer equations for compressible flow are not significantly more complex than those for incompressible flow. The main difference is in the property variations of p, k , and p. The boundary layer approximation can also be made for a threedimensional flow as long as velocity derivatives with respect to only one coordinate direction are large. Thus, the three-dimensional boundary layer flow remains "thin" with respect to only one coordinate direction. With y-direction normal to the wall, the three-dimensional boundary layer equations in Cartesian coordinates, applicable to a compressible flow, are continuity : (1.43) momentum : au au au p j p p u - +ax p v - + payw - =

ap -ax+$(p$)

(1.44) (1.45)

energy :

aH ay

pu - 4- pw -

2 ay [p{L Pr ay + (1

I})$

(1.46)

The equation of state is the same as Eq. (1.42),and the enthalpy H is UL

H=c,T+-+

WL 2

The three-dimensional boundary layer equations are used primarily for external flows, for which the pressure gradient terms can be evaluated from the solution of the Euler equations. Three-dimensional internal flows are

Garg

generally computed from somewhat different equations, to be discussed in section 1.10. Also, unlike the Navier-Stokes equations, the boundary layer equations are purabolic in the main flow direction, and thus the solution can be "marched" in that direction. This makes boundary layer equations relatively easier to solve than the ellipric Navier-Stokes equations.

1.9

EULER EQUATIONS

We observed in the preceding section that for solution of the boundary layer equations, we must first solve for the inviscid flow outside the boundary layer. Note that the inviscid part of the flow can be solved independently of the boundary layer part only if the boundary layer is very thin compared to a characteristic length of the flow field, so the interaction between the two parts is negligible. For flows in which this interaction is not negligible, it is still possible to use separate sets of equations for the two regions, but the equations must be solved iteratively. This iterative procedure can be computationally inefficient. It is then desirable to use a single set of equations that remain valid throughout the flow field. Such equations will be discussed in section 1.10. In this section, we discuss equations that are valid only in the inciscid (nonviscous, nonconducting) portion of the flow field. These equations are obtained simply by dropping the viscous and heat transfer terms from the Navier-Stokes equations. As a consequence of this reduction, inviscid flow equations are much simpler to solve than the Navier-Stokes equations. Some of these simplifications will be observed here. Strictly speaking, Euler equation refers only to the inviscid momentum equation. However, we will refer to the set of inviscid flow equations as the E u k r equations. 1.9.1

Continuity Equation

The continuity equation does not contain any viscous or heat transfer terms. Thus simplification of the continuity equation for an inviscid flow is not possible. However, for a two-dimensional or axisymmetric, steady flow, it is possible to satisfy the continuity equation exactly by introducing a streamfunction II/. This is true irrespective of the flow being viscous or inviscid. For example, the continuity equation for a two-dimensional, steady, compressible flow in Cartesian coordinates is ( 1.47)

Governing Equations

If we define a stream function t j such that

pu = -, ay

pu=--

( 1.48)

it is clear that Eq. (1.47) is satisfied. Thus introduction of the stream function reduces the number of dependent variables by one. The price for this reduction is that the velocity derivatives in the remaining equations have to be replaced using Eq. (1.48). These remaining equations will therefore contain derivatives which are one order higher. The physical significance of the stream function is obvious from

=pV*dA=dh

(1.49)

Thus there is no mass flow ( d h = 0) across lines of constant $ (d$ = 0). Lines of constant $ are called streamlines, and the difference between the values of $ for any two streamlines represents the mass flow rate per unit width between those streamlines. By definition, a streamline is a line in the flow field whose tangent at any point is in the direction of the flow velocity at that point. For a two-dimensional, incompressible flow the continuity equation in Cartesian coordinates is

-a +u - =aou ax ay

(1 S O )

and the stream function is defined by U=-

all/ , ay

U=--

all/ ax

(1.51)

For a steady, axisymmetric, compressible flow the continuity equation in cylindrical coordinates is (see Appendix A) (1.52)

and the stream function is defined by 1pl/ = -

r dz’

a$ p v , = - -1 r dr

(1.53)

It is possible to replace the continuity equation for three-dimensional flows by two stream functions. However, it is more complex than using the continuity equation in its original form and is therefore seldom used.

Garg

1.9.2

lnviscid Momentum Equations

The inviscid momentum equation, also known as the Eufer equarion, is given by Eq. (1.14). If we assume steady flow and neglect body forces, i t reduces to

v * v v = - - v1 p

(1.54)

Integrating this equation along a line in the flow field gives ( V ’ vv) dr

= -

v p . dr

(1.55)

where dv is the differential length along the line. Let us further assume that the line is a streamline. Thus, V has the same direction as dr and we can simplify the integrand on the left side of Eq. (1.55) to get

Similarly, the integrand on the right side of Eq. (1.55) yields 1 dP Vp dr = P r

and Eq. (1.55) reduces to

E + 2

= constant

(1.56)

The integral in Eq. (1.56) can be evaluated if the flow is assumed barotropic, i.e, a fluid for which p is either constant or a function only of p ; the former being an incompressible flow, and the latter an isentropic flow for which p = (constant) p l ’ ~

( 1.57)

Thus for a steady incompressible flow, Eq. (1.56) reduces to p

+ -21 p V 2 = constant

(1.58)

which is called the Bernoulli equation. For a steady isentropic, compressible flow, Eq. (1.56) yields

-V -2 + Y E -

(1.59) - constant 2 Y-1P which is sometimes referred to as the compressible Bernoufli equation. Note that Eqs. (1.58) and (1.59) are valid onlj, along a y i w n streamline.

Governing Equations

It can be easily shown that Eqs. (1.58) and (1.59) are valid everywhere in the flow field if the flow is irrotational as well. For an irrotational flow the vorticity is zero; i.e., the fluid particles do not rotate about their axis. The vorticity & is defined by < = v x

(1.60)

Thus, for an irrotational flow, V can be expressed as the gradient of a single-valued scalar 4 since & = V x V=Vx(V+)=O

(1.61)

The scalar 4 is called the velocity potential. Also, we can express the acceleration of a fluid particle as D V- - av + Dt at

v.vv=~+v(;)-vx& at

(1.62)

For an irrotational flow this equation reduces to DV Dt

av at

which can be substituted into Euler’s equation [Eq. (1.14)] to yield (1.63) If we again assume steady flow and neglect body forces, Eq. (1.63) can be written as

v(;

):1 +

(1.64)

since

Integrating Eq. (1.64) along any arbitrary line in the flow field yields Eq. (1.56) again, but the constant now has the same value everywhere in the flow field. The incompressible Bernoulli equation [Eq. (l.S8)] and the compressible Bernoulli equation [Eq. (1.59)] follow in the same manner as before, the only difference being that the equations are valid everywhere in the flow field because of the additional assumption of irrotationality. For the special case of an inviscid, incompressible, irrotational flow, the continuity equation

v*v=o

Garg

can be combined with V=V$ to yield the Laplace equation

VZ$

lnviscid Energy Equations

1.9.3

We have already seen that for an iidiahutic flow of an inviscid fluid, Eq. (1.21) holds. This can also be written in terms of specific enthalpy as Dh

PE=-Dt

( 1.66)

Defining the specific stagnation enthalpy H as VZ H = ~ + T we can rewrite Eq. ( I .66) as DH (7p p--=7+Vg Df

( I .67)

c t t

If the body forces are neglected, the term within parentheses on the right side vanishes due to the Euler equation [Eq. (1.14)]. Equation (1.67) thus reduces to

D H - 1 (7p Dt p (7t

(1.68)

which for a steady flow becomes V*VH=O This equation can be integrated along a streamline to yield H For an incompressible flow, Eq. (1.21) reduces to

-Df? =o Dt

( 1.69) =

constant. ( 1.70)

which implies that the internal energy is constant along a streamline in steady flow. 1.9.4

Simplified Form of Euler Equations

I t is possible to simplify the Euler equations by making additional assumptions. If we assume the flow to be steady, irrotational, and isentropic, the Euler equations can be combined into a single equation called the w l o c i f y p o t t w i d eqciiirion. In the Cartesian coordinate system, replacing the veloc-

Governing Equations

ity components by U = - 84 2X

U=-

dY '

w = - a4

the continuity equation can be written as (1.71)

where the subscript denotes partial differentiation with respect to the variable. The momentum and energy equations reduce to Eq. (1.56) under the present assumptions. In differential form this equation can be written as ( 1.72)

With the speed of sound defined by

Eq. (1.72) can be written as ( 1.73)

which can be used to find the derivatives of p in each direction. Substituting these expressions for p x , p y , and pz into Eq. (1.71) and simplifying, the velocity potential equation is obtained: (1 - $)4,.

+ (1

- $)4yy

+ (1

- $)L

For an incompressible flow, the velocity potential equation reduces to the Laplace equation as a -+ CO. 1.10

PARABOLIZED NAVIER-STOKES EQUATIONS

The boundary layer equations can be used to solve many, but not all, viscous flow problems since the boundary layer assumptions are invalid for some viscous flow problems. For example, if the inviscid flow is fully merged with the viscous flow, the two flows cannot be solved independent of each other as required by boundary layer theory. It then becomes necessary to solve a set of equations which are valid in both the inviscid and

Garg

viscous flow regions. Some examples of viscous flow fields where the boundary layer equations are not appropriate include (1) a supersonic flow around a blunt body at high altitude, and (2) a flow along a corner formed by two intersecting surfaces, among others. In the first example, there is a strong interaction between the boundary layer and inviscid flow in the region between the shock wave and the blunt body, while in the second example, viscous derivatives with respect to two “normal” directions are important very near the corner. As we pointed out earlier, the boundary layer equations include viscous derivatives with respect to a single ‘‘normal ” coordinate direction only. Obviously, the complete NavierStokes equations can be used to solve such flow fields. However, they are very difficult to solve. In general, a large amount of computer time and storage is required to solve the complete Navier-Stokes equations. Fortunately, for some of the viscous flow problems for which the boundary layer equations are not appropriate, it is possible to solve a reduced set of equations that fall between the complete Navier-Stokes equations and the boundary layer equations in terms of complexity. These reduced equations are often referred to as the “parabolized” Navier-Stokes equations. They contain a nonzero normal pressure gradient which is a necessary condition for solving the viscous and inviscid flow regions simultaneously. The most important advantage in using the parabolized Navier-Stokes (PNS) equations instead of the complete Navier-Stokes equations is that for a steady flow the former are a mixed set of hyperbolic-parabolic equations in the streamwise direction, provided certain conditions are met. In other words, the Navier-Stokes are “parabolized” in the streamwise direction, leading to a boundary-layer-type marching technique for the solution. This results in a substantial saving in computer time and storage. Another saving in computation time results from the fact that the PNS equations have fewer terms compared to those in the complete equations. The conditions under which the PNS equations are a set of hyperbolic-parabolic equations are that the inviscid outer region of the flow be supersonic, and that the streamwise velocity component be positive everywhere. The last condition excludes streamwise flow separation but crossflow separation is permitted. An additional complication is caused by the presence of streamwise pressure gradient in the streamwise momentum equation. With this term present everywhere in the flow field, upstream influence can occur in the subsonic part of the boundary layer, and a space-marching technique is not well-posed. The reader is referred to Anderson et al. (1984) for techniques to overcome this difficulty. Different versions of the PNS equations are available in the literature since their derivation from the complete Navier-Stokes equations is not as rigorous as that of the boundary layer equations. These versions differ

Governing Equations

sometimes based on the type of flow problem being solved. In all cases, however, the normal pressure gradient term is retained, and the second derivative terms with respect to the streamwise direction are omitted. Rudman and Rubin (1968) were perhaps the first to use the PNS equations to study the hypersonic laminar flow near the leading edge of a flat plate. They used a series expansion technique to reduce the complete NavierStokes equations to the PNS equations. The set of PNS equations derived by Kudman and Rubin does not contain a streamwise pressure gradient term. Thus, no upstream influence through the subsonic part of the boundary layer is allowed, and the equations behave in a strictly parabolic manner in the boundary layer region. These PNS equations have been used to solve leading-edge flows about both two- and three-dimensional geometries including flat plates, rectangular corners, cones, and wing tips (see Lin and Rubin, 1973, for references). Leaving the details to the reader (see, for example, Anderson et al., 1984), the three-dimensional PNS equations of Rudman and Rubin in Cartesian coordinates are continuity : at

(1.75)

x-momen t um : a u -f pv

au + pw au = a (p$)+;(Pg)

(1.76)

y-momentum : p

av au + pu av + p v -ay +pw aZ

(1.77)

z-momentum : p

aw + pu aw aW aw + pv ay + pw ax

= _ -ap

$) + ;; (. 2) E)+ -$ ); ; $ + g) a2

+ (3 ( P

+ ax ( P

(1.78)

Garg

energy:

pc" -

?T c:T ?T + puc,, + p"[> + p\vc, (7 s ?y (7:

( 1.79)

The most common form, perhaps, of the PNS equations (Lubard and Helliwell, 1973, 1974) is obtained by assuming that the streamwise viscous derivative terms (including the heat flux terms) are negligible [assumed to be of O( l)] compared to the normal and transverse viscous derivative terms [of O(Rei, ')I. The resulting set of equations for a Cartesian coordinate system (sis the streamwise direction) is continuity : ( 1.80)

.\^-momentum:

j'-momentum :

:-momentum :

(

"':>

+ 7 P -(7z - OJ* 3 (72

( i:>

P 7,

( 1.83)

Governing Equations

energy : i?T dT 8T 8T + puc, + pvc, + pwc, at ax ay aZ

pc, -

(1.84)

A comparison of this set of PNS equation with that of Rudman and Rubin [Eqs. (1.75) to (1.79)] reveals that while the continuity and energy equations are identical, the momentum equations are different. In particular, the present x-momentum equation contains the streamwise pressure gradient term.

1.11 GOVERNING EQUATIONS IN GENERALIZED COORDINATES

For many problems, a transformation from physical space to computational space is performed in order to simplify implementation of the boundary conditions and to enhance the eficiency and accuracy of the numerical scheme. This transformation allows clustering of grid points in regions where the flow variables undergo high gradients, and grid point motion when required. The computational domain is rectangular with a uniformly spaced grid. While grid generation is discussed in Chapter 4, it is clear that a transformation of the governing equations from the physical space into the computational space is also required before any solution can be obtained. In this section we will see how the governing equations can be transformed from a Cartesian coordinate system (x, y, z) in the physical space to any general nonorthogonal (or orthogonal) coordinate system (t,q, in the computational space. The governing equations are written in strong conservative form to include the capability for shock capturing (Anderson, 1992a). Let us consider a completely general transformation of the form

Garg

The chain rule of partial differentiation yields the following for the Cartesian derivatives:

(13 6 )

r,,

The metrics (51, Z l t , Cl, t,, q x , 5 p q y 7 i y , 72, CZ) appearing in these equations must be evaluated. In most cases analytical determination of the metrics is not possible; therefore they must be computed numerically. Since the grid size in the computational space is uniform, x g , x,, xY,etc.. can be computed readily. Thus, if the metrics appearing in Eqs. (1.86) can be expressed in terms of these derivatives, the numerical computation of metrics is easy. In order to obtain such relations, we first write the differential expression 5 2 7

( 1.87)

dt = dz

Similarly dx

+ X< d t + X, dq + X; dc dz + yc d y + yrl dq + y; d( dz + zi dc + Z, dq + zZ dC

= .U, d~

dy =

dz = Z ,

( 1.88)

Equations (1.87) and (1.88) can be expressed in matrix form as 0

Y, . )

(139)

Governing Equations

Reversing the role of the independent variables, we can write (1.90)

Comparing Eqs. (1.89) and (1.90), we get

-1

1 0 0 0 -

5, 5, 5,

flt

‘Ix

fly

flz

52-

(1.91)

Thus, the metrics are

and

flz

(1.95)

which can be evaluated as follows:

(1.96)

Garg

For a discussion of the proper way to compute metrics, see Anderson et al. 10).Applying the generalized transformation to the compressible Navier-Stokes equations written in vector form (Eqs. ( 1.27)], we obtain the transformed equation

( 1984, Chap.

U , + t, U: + qr I/tl + i t U; + 5, E: + E, + i., E; ( 1.97) + t,.F; q , F q [,.F; t z G : qzCq+ CzG; = 0 This equation is no longer in conservative form. Following Viviand (1974) and Vinokur (1974), i t can be cast into strony c'onsrriwtiilc, form to yield

( 1.98)

We can redefine the terms in this equation to write it in the form ( 1.99)

where

(1.100)

Note that the vectors E , , F , , and G1 contain viscous and heat flux terms [cf. Eqs. (1.28)] that involve partial derivatives. These partial derivatives are also to be transformed using Eqs. (1.86). The viscous stresses, using Stokes' hypothesis (p' = -2p,'3), in the transformed computational space are

Governing Equations

+ c, U[) - 32 (5, U < + l', + i,";I zxy = ,z, = P(t,U< + 'I,,ua + c y U [ + 5, + + i,t'J 7,- = 72, = P ( L U < + vl* U, + c, U[ + w< + M', + ";I z,,,= z,, = p(5, U < + q, 0, + c, V [ + 5, w< + vlr "', + i,1 q -

52 (tXU < +

vly

'Ix t'q

'Ix

(1.101)

i x

and the heat conduction terms in the computational space are (1.102) The conservative form of the governing equations is convenient for applying finite-difference schemes (Anderson, 1992a). For many applications, the grid is independent of time, and thus time gradients of the metrics are zero. However, when the grid is changing with time, a constraint on the way the metrics are differenced, called the geometric conseriwtion law (Thomas and Lombard, 1978), must be satisfied so as to prevent the introduction of additional errors into the solution. Details are available in Anderson et al. (1984). 1.12 MATHEMATICAL PROPERTIES OF THE GOVERNING EQUATIONS

It is useful to examine some mathematical properties of the Navier-Stokes equations since any valid solution of the equations should obey these properties. Depending upon the flow situation, the Navier-Stokes equations can be classified as hyperbolic, elliptic, or parabolic. For the rules governing this classification, the reader may refer to any text on partial differential equations or to Anderson et al. (1984), Fletcher (199 1, Chap. 2). and Anderson (1992b). We make some relevant observations here. For hjiperbolic equations, information at a given point influences only those regions between the advancing characteristics. The method of d i ~ r a c teristics takes advantage of this property during solution; one can march

Garg

along the characteristic. For parabolic equations, information at a point P in, say, the X-J. plane influences the entire region of the plane to one side of P only. The solution to parabolic equations can thus be marched in the main flow direction. For elliptic equations, information at a point P in the x-y plane influences all other regions of the domain. Therefore, the solution at point P must be carried out simultaneously with the solution at all other points in the domain. This is in stark contrast to the ‘marching’ solutions germane to parabolic and hyperbolic equations. For inciscid compressible flows, the system of equations is always hyperbolic- if the motion is unsteady, no matter whether the flow is locally subsonic or supersonic. If the motion is steady, the classification of the system depends upon the fluid speed, being hyperbolic if it is supersonic and elliptic if subsonic. Thus, inviscid compressible flows can be of mixed type when the flow is steady and subsonic in one region while being supersonic elsewhere. Since incompressible flow (which theoretically implies that the Mach number is zero) is a subcase of the subsonic flow, steady inuiscid incompressible flows are also elliptic. For such flows, physical boundary conditions must be applied over a closed boundary that totally surrounds the flow domain. Examples of parabolic flows are the boundary layer flows governed by the boundary layer equations of section 1.8 and flows governed by the PNS equations of section 1.10. Owing to the totally different mathematical behavior of elliptic and hyperbolic equations, we can appreciate the difficulties early researchers encountered in trying to solve mixed problems, a prime example of which is the supersonic flow over a blunt body such as an atmospheric entry vehicle. The sudden change in the nature of the governing equations across the sonic line precluded any practical solution of the steady flow blunt body problem involving a uniform treatment of both the subsonic and supersonic regions. However, since unsteady inviscid flow is governed by hyperbolic equations, no matter whether the flow is locally subsonic or supersonic, we can solve the unsteady inviscid flow equations, marching forward in time. At large times, the solution approaches steady state, which is the desired result. Thus, we get the steady-state solution for the entire flow field, including both the subsonic and supersonic regions, using the same, uniform method throughout the flow field. This is the basic philosophy behind the time-dependent technique for the solution of flow problems. It may be helpful to the reader to examine the closed-form solution to some linear partial differential equations of the elliptic, parabolic, and hyperbolic types. Numerous classical solutions can be found in Hildebrand (1976) and Anderson et al. (1984). We move on to numerical solutions in the remainder of the book.

Governing Equations

NOMENCLATURE

specific heat specific heat at constant pressure specific heat at constant volume specific internal energy total energy per unit volume body force per unit volume specific ent halpy ent halpy Jacobian of the coordinate transformation thermal conductivity characteristic length mass flow Mach number pressure surface force per unit volume Prandtl number conductive heat flux gas constant Reynolds number time temperature Cartesian velocity components velocity vector radial and axial components of velocity in cylindrical coordinates Cartesian coordinates

Greek

thermal diffusivity coefficient of volumetric expansion ratio of specific heats ( = cp/c,) Kronecker delta ( = 0 for i # j , and vorticity coefficient of bulk viscosity dynamic viscosity second coefficient of viscosity kinematic viscosity generalized coordinates

1 for i = j )

P gij

cf)

a) z zij

Garg

density stress on a plane normal to axis i in the direction of axisj velocity potential dissipation function time viscous (shear) stress tensor stream function

Subscripts Y

-L

refers to value at the edge of the boundary layer refers to value in the freestream

REFERENCES Anderson DA, Tannehill JC, Pletcher RH Computational Fluid Mechanics and Heat Transfer. Washington DC : Hemisphere: 1984. Anderson J D Jr. Governing equations of Auid dynamics. In: Wendt JF, ed. Computational Fluid Dynamics An Introduction. Berlin: Springer, 1992a pp 15 5 I . Anderson J D Jr. Mathematical properties of the fluid dynamic equations. In: Wendt J F. ed. Computational Fluid Dynamics A n Introduction, Berlin: Springer, 1991b, pp 75-84. Cebeci T, Smith AMO. Analysis of Turbulent Boundary Layers. New York: Academic Press; 1974. E'letcher CAJ Computational Techniques for Fluid Dynamics. 2nd ed. Berlin: Springer, 1991. Hildebrand FB. Advanced Calculus for Applications. Englewood Cliffs, NJ : Prentice-Hall, 1976. Lin TC, Rubin SG Viscous flow over a cone at moderate incidence. I : Hypersonic tip region. Comput Fluids 1 : 37 -57, 1973. Lubard SC. Helliwell WS. Calculation of the flow on a cone at high angle o f attack. R&D Associates Technical Report, RDA-TR- 150, Santa Monica. CA, 1973. Lubard SC,Helliwell WS. Calculation of the flow on a cone at high angle of attack. A I A A J 12: 965-974, 1974. Prandtl L. Verh. 3rd Intl. Math. Kongr. Heidel, p 484 (translated as NACA T M 452), 1984. Round G F , Garg V K . Applications of Fluid Dynamics. London: Edward Arnold. 1986. Rudman S, Rubin SG Hypersonic viscous flow over slender bodies with sharp leading edges AIAA J 6: 1883 - 1889, 1968. Schlichting H. Boundary Layer Theory. 7th ed. New York: McGraw-Hill, 1979. Thomas PD, Lombard CK The geometric conservation law -a link between finitedifl'erence and finite-volume methods of flow computation on moving grids. A I A A Paper 78-1208, Seattle, WA, 1978.

Governing Equations

Vinokur M. Conservation equations of gas-dynamics in curvilinear coordinate systems. J Comp Phys 14: 105-125, 1974. Viviand H. Conservation forms of gas dynamic equations. La Recherche Aerospatiale, No 1974-1, pp 65-68, 1974.

This page intentionally left blank

Numerical Tec hniq ues Vijay K . Garg A Y T CorporationlNASA Lewis Research Center. Cleoeland. Ohio

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . 2.2 Finite Difference Method . . . . . . . . . . . . . . . 2.2.1 Finite Difference Representation of Derivatives . . . . . 2.2.2 Finite Difference Approximationsat aBoundary . . . . . 2.2.3 Solution of Finite Difference Equations-Some Considerations 2.2.4 Errors and Stability Analysis . . . . . . . . . . . . 2.2.5 Convergence . . . . . . . . . . . . . . . . . 2.3 Finite Element Method . . . . . . . . . . . . . . . . 2.3.1 Weighted Residual Formulation . . . . . . . . . . . 2.3.2 Galerkin Formulation . . . . . . . . . . . . . . 2.3.3 Application to an Element . . . . . . . . . . . . . 2.3.4 Assembly . . . . . . . . . . . . . . . . . . . 2.3.5 Elements . . . . . . . . . . . . . . . . . . . . 2.3.6 Condensation and Substructuring . . . . . . . . . . 2.4 Finite Volume Method . . . . . . . . . . . . . . . . 2.4.1 Basic Considerations . . . . . . . . . . . . . . . 2.4.2 Two-Dimensional Finite Volume Method . . . . . . . 2.4.3 Integration Formulas for Finite Volumes . . . . . . . . 2.4.4 Three-Dimensional Finite Volume Method . . . . . . . 2.4.5 Time Integration . . . . . . . . . . . . . . . .

. . . . .

36 37 38 . . 41 . . 42 . . 44 . . 48 . . 49 . . 50 . . 51 . . 52 . . 53 . 54 . . 56 . . 56 . . 59 . . 61 . . 64 . . 65 . . 68 35

2.1

Garg

Nomenclature . . . . . . . . . . . . . . . . . . . . .

References

. . . . . . . . . . . . . . . . . . . . . .

INTRODUCTION

We notice from Chapter 1 that the equations governing fluid flow and heat transfer are complex partial differential equations for which no analytical solution can be found except in rather simple situations. For many practical problems, in general, obtaining a numerical solution to the Navier-Stokes equations is the only possibility, short of actual experimentation. With the easy availability of high-speed digital computers, growth of computational fluid dynamics (CFD) has mushroomed in the last three decades. While analytical solutions, if possible, yield closed-form expressions which give the variation of dependent variables c-ontinuouslj~throughout the domain, numerical solutions can give answers only at discrete points in the domain, called grid points. The end product of C F D is thus a collection of numbers, in contrast to a closed-form analytical solution. In the long run, however, the objective of most engineering analyses, closed-form or otherwise, is a quantitative description of the problem, i.e., numbers (Anderson, 1976). For incompressible flows there are several possibilities for the formulation of the problem. These include primitive variables, stream-function vorticity, and vorticity velocity formulations. The primitive variable approach offers the fewest complications in extending two-dimensional schemes to three dimensions. The primary difficulty with this approach is the specification of boundary conditions on pressure. See Peyret and Taylor (1983) and Anderson et al. (1984) for ways to overcome this problem. The streamfunction vorticity formulation is best suited for two-dimensional flows, though it has been applied to three-dimensional incompressible flows as well [see, for example, Aziz and Hellums (1967), and Mallinson and De Vahl Davis (1973, 1977) for application details, and Hirasaki and Hellums ( 1970) and Richardson and Cornish (1977) for boundary condition considerations]. The difficulty with such methods is primarily associated with determination of vorticity at a boundary. A number of ways to overcome this difficulty are available (Peyret and Taylor, 1983; Anderson et al., 1984). An inconvenience of this formulation is that the pressure is not directly available, and additional computation is required for its determination. We may point out that for two-dimensional flows, the streamfunction-only formulation can also be utilized. The interested reader is referred to Bourcier and Franqois (1969), Roache and Ellis (1975). Morchoisne ( 1979), Cebeci et al. (1981), and Jaluria and Torrance (1986)

Numerical Techniques

for the details. The vorticity velocity formulation requires the vorticity equation, the continuity equation, and the equations that define vorticity in terms of velocity gradients. A combination of the continuity equation and the definition of vorticity yields elliptic equations for the velocity components. The interested reader is referred to Fasel (1976) and Dennis et al. (1979) for details. Today there are a number of numerical techniques available for solving the fluid flow and heat transfer problems. Some of these are specific to the type of flow under investigation, as mentioned above, and for example, the panel methods for inviscid incompressible flows, Green’s function methods for incompressible flows (generally two-dimensional), the method of characteristics for inviscid supersonic or inviscid transient compressible flow problems, etc. In this chapter, however, we will describe briefly three techniques that, in principle, encompass all the methods. These are the finite difference method (FDM), the finite element method (FEM), and the finite volume method (FVM). Peyret and Taylor (1983) point out that the spectral method can be considered as a variant of the FEM. For some relationships between the FDM, FEM and spectral methods, see Patankar (1980), Peyret and Taylor (1983), and Fletcher (1991). The relative merits of the FDM, FEM, and spectral methods are given in Fletcher (1984, Chap. 6). 2.2

FINITE DIFFERENCE METHOD

The finite difference method is widely used and is perhaps the oldest method. The essence of a finite difference method is to replace the partial derivatives appearing in the governing equations with algebraic difference quotients, thus yielding a system of algebraic equations which can be solved for the flow-field variables at the specific, discrete grid points in the flow domain. The nature of the resulting algebraic system depends upon the character of the problem defined by the original partial differential equation (PDE) or system of PDEs. For convenience, let us consider a two-dimensional problem, and let Fig. 2.1 show a section of the discrete grid in the x-y plane. Let us assume that the spacing of the grid points in the x-direction is uniform, and given by Ax, and that the spacing of grid points in the y-direction is also uniform, and given by Ay, as shown in Fig. 2.1. In general, Ax and A J ~are different. Indeed, it is not necessary that Ax and Ay be uniform; we can have totally unequal spacing in both directions so that Ax is a different value between each successive pair of grid points, and similarly for Ay. The vast majority of C F D applications, however, involve numerical solutions on a grid which involves uniform spacing in each direction, since this greatly simplifies the

Garg

Figure 2.1

Finite difference grid in a two-dimensional region.

programming, saves storage space, and usually results in greater accuracy. This uniform spacing does not occur in the physical x-y space. Generally, in CFD, numerical calculations are carried out in a transformed computational space which has uniform spacing in the transformed independent variables, but which corresponds to nonuniform spacing in the physical space. The grid lines are indexed by integers i a n d j which increase monotonically along the x and 4’ coordinates, respectively. Let us now derive some common finite difference expressions used to replace the partial derivatives in the PDEs. 2.2.1

Finite Difference Representation of Derivatives

Finite difference representation of the partial derivatives can be derived from Taylor series expansion. Let h, denote the value of dependent variable f at the point (i, j) with coordinates ( x i , yj), where f may be a component of fluid velocity, temperature or any other dependent variable. Then the value fi+ 1 , at point (i + 1, j ) can be expressed in terms of the Taylor

Numerical Techniques

series expanded about point ( i , j ) as follows:

Equation (2.1) is a mathematically exact expression for fi+ if (a) the number of terms is infinite and the series converges, and/or (b) Ax -+ 0. Clearly, only a finite number of terms in Eq. (2.1) can be carried in numerical computations. Thus, Eq. (2.1) is truncated. For example, if terms of order ( A x ) ~and higher are neglected, Eq. (2.1) reduces to

Equation (2.2) is said to befirst-order accurate. If terms of order ( A x ) ~and higher are neglected, we get from Eq. (2.1)

where Eq. (2.3) is second-order accurate. In Eqs. (2.2) and (2.3), the neglected higher-order terms represent the truncation error in the finite series representation. The truncation error can be reduced by (a) carrying more terms in the Taylor series expansion, Eq. (2.1), leading to higher-order accuracy in the representation of fi,j , and (b) reducing the magnitude of Ax. We can solve Eq. (2.1) for (aflax),, to get

or & + l ,j

-& + O(Ax)

The symbol O(Ax) in Eq. (2.4) is a formal mathematical notation which represents “terms of order of Ax.” Details on the 0 notation can be found in Whittaker and Watson (1927). Equation (2.4) is a more precise notation than Eq. (2.2), which involves the “approximately equal” sign. The firstorder-accurate difference representation for the derivative (afldx),, j , expressed in Eq. (2.4), is known as the first-order forward diflerence. Similarly, the first-order backward diflerence expression for the derivative (afldx),, can be written as

Garg

which follows from the Taylor series expansion f o r f , - , , about the point

(i.j),obtained by replacing A s by -A.Y on the right side of Eq. (2.1) to yield

(2.6)

Subtracting Eq. (2.6) from Eq. (2.1), and rearranging, we get

which is the second-order centrul diference for the derivative (?&?X)~, To obtain a finite difference expression for the second partial derivative ( ? ? / ' ? . Y ~ ) ~j , , we add Eqs. (2.1) and (2.6) and rearrange to yield (2.8) Equation (2.8) is a second-order central second dtference for the derivative (il:f'/(7s2) at grid point ( i , j ) . Difference expressions for the ),-derivatives are obtained in the same manner. The results, analogous to the previous equations for the sderivatives, are Forward difference

(g)

j: . -j: J,

. I*

J-

Backward difference

O(Ay)

Central difference

.fi. j +

- 2L.j

+fl,

j- 1

+ O(AJ)~

Central second difference

We leave i t as an exercise for the reader to show that the central second difference in Eq. (2.8) can be interpreted as a forward difference of the first derivatives, with backward differences used for each of the first derivatives. A similar philosophy can be used to generate a finite difference expression for the mixed derivative (?2f[?x?~*) at grid point ( i , j ) . Clearly. (2.9)

Numerical Techniques

In Eq. (2.9), we write the x-derivative as a central difference of the yderivatives, and then write the y-derivatives also in terms of central differences to get

Many other difference approximations can be written for the above derivatives, as well as for derivatives of higher order. For a tabulation of many forms of difference approximations, see Hyman and Larrouturou (1982). 2.2.2

Finite Difference Approximations at a Boundary

At a boundary, only one direction, away from the boundary, is available. For example, Fig. 2.2 illustrates a portion of the boundary with a normal in the y-direction. Let grid point 1 be on the boundary, with points 2 and 3 a

-1. It2

Figure 2.2

Boundaw

Grid points at a boundary.

Garg

distance Ay and 2 Ay above the boundary, respectively. It is easy to write a forward difference approximation for ?fl?j?at the boundary as

(g)

f2-fi + O(Ay) AY

(2.11 j

which is first-order accurate. The second-order accurate approximation for ?j/l‘y at the boundary can be found by fitting a parabola f=a+

bq’+cy2

(2.12)

through the three points 1, 2, and 3 in Fig. 2.2. This yields

(2.13) Using a Taylor series expansion about point 1, it can be shown that Eq. (2.13) is second-order accurate. Thus, the second-order-accurate approximation for ifl?y at the boundary is

(2.14) Both Eqs. (2.11 ) and (2.14) are called one-sided di$erences, since they express a derivative at a point in terms of dependent variables on only one side of the point. Many other one-sided differences can be formed, with higher degrees of accuracy, using additional grid points (Hyman and Larrouturou, 1982). 2.2.3

Solution of Finite Difference Equations---Some Cons ider at ions

As stated earlier, the FDM replaces the partial derivatives in the governing equations by the difference quotients, leading to a system of algebraic equations for the dependent variable(s) at each grid point. Let us examine some aspects in the solution of these equations by considering the model equation

(2.15) Here, we have assumed, for convenience, that the dependent variable f’ is a function of .Y and t only. There is no advantage at this stage in considering a more complex example. It may be noted that Eq. (2.15) is parabolic, and its solution can therefore be “marched” (cf. section 1.12). Replacing the time derivative in Eq. (2.15)with a forward difference, and the spatial derivative

Numerical Techniques

with a central difference at time t, we get (2.16) In Eq. (2.16), the index for time appears as a superscript, where n denotes values at time t, n + 1 denotes values at time t + At, and so on. The subscript denotes the grid point location, as usual. Since forward difference is used for the time derivative, and central difference for the spatial derivative, the truncation error for the complete equation is O[At, (Ax)']. If truncation error approaches zero as the grid is refined (in general, Ax, Ay, Az, and At + 0), the finite difference approximation is said to be consistent with the governing differential equation. Consistency is an essential requirement for a valid numerical simulation of a partial differential equation. In our example, we note that the truncation error approaches zero as A x - 0 and A t - 0 . Hence the difference Eq. (2.16) is consistent with the differential Eq. (2.15). Let us now examine the solution of Eq. (2.16). From the given initial condition, we know the dependent variable at all x at the initial instant. Examining Eq. (2.16), we find that it contains only one unknown, namely fl". Thus, the solution of Eq. (2.16) can be marched in time such that the value off at time t + At can be obtained explicitly from the known values at time t ; i.e., fl+' is obtained directly from the known valuesff+l, fl, and fl- '. This is an example of an explicit Jinite diflerence solution. As a counterexample, let us replace the spatial derivative in Eq. (2.15) by the central difference at time t + At, instead of at time t as earlier, to get (2.17) The differencing given in Eq. (2.17) is the fully implicit form. If the spatial difference on the right side of Eq. (2.17) is written in terms of the average values between time t and t + At, we get the well-known Crank-Nicolson form. Examining Eq. (2.17), we find that the unknown is not only expressed in terms of the known values at time index n, namelyfl, but also in terms of unknown values at time index n + 1, namely fl;; and fl;. Hence, Eq. (2.17) applied at a given grid point i cannot by itself result in the solution for fl". Rather, Eq. (2.17) must be written at all grid points, resulting in a system of algebraic equations, which must be solved simultaneously to yield the unknownfl" for all i. This is an example of an implicit Jinite diflerence solution. Note that the system of algebraic equations that results in sparse. For the present example, we get a tridiagonal system of equations, which can be easily solved using an algorithm based on Gaussian elimination, often known as the Thomas algorithm (Hornbeck, 1973).

Garg

The number of arithmetic operations required for solving a tridiagonal system of tz equations by this algorithm is O(n), as compared to O ( i i 3 ) needed for the solution of a general algebraic system by Gaussian elimination. This leads to a very efficient algorithm and small roundqfi’error. The relative major advantages and disadvantages of the explicit and implicit methods are : Esplicit Approach Advantage: relatively simple to set up and program. b. Disadvantage: For a given spatial grid, At must be less than some limit imposed by stability constraints. In many cases, Ar must be very small to maintain stability, resulting in a long computational time to obtain results over a given interval of time. 2. Implicit Approach Advantage: stability can be maintained over much larger values of At, thereby requiring fewer time steps to obtain results over a given interval of time. For the fully implicit approach, there is no stability constraint on At. Note, however, that the Crunk-Nicolson scheme, contrary to popular belief, can lead to physically unrealistic solutions (Patankar and Baliga, 1978; Patankar, 1980). Disadvantage: more complicated to set up and program. Disadvantage : since matrix manipulations are usually required at each time step, computer time per time step is much larger than that in the explicit approach. Disadvantage: since large At can be taken, truncation error is larger. Thus, use of the implicit method to follow the exact transients (time variations of the dependent variable) may not be as accurate as an explicit approach. However, if the steady state is the desired result, the relative timewise inaccuracy for a time-dependent solution is unimportant. 1.

2.2.4

Errors and Stability Analysis

While the FDM is a philosophically straightforward technique, there is no guarantee that the solution of the algebraic equations will be accurate, or even stable, under all conditions for a given problem. For linear PDEs, however, there is a formal way of examining the accuracy and stability, and these ideas provide some guidance for understanding the behavior of more complex nonlinear PDEs that govern fluid flow and heat transfer, in general. While more details are available in Hornbeck (1973) and in Ander-

Numerical Techniques

son et al. (1984), we present some basic ideas here. The numerical solution of a P D E is influenced by two sources of error: 1. Discretization Error: The difference between the exact analytical

solution of the PDE [for example, Eq. (2.15)] and the exact solution (to infinite precision) of the corresponding difference equation [for example, Eq. (2.16)]. Based on the previous discussion, the discretization error is simply the truncation error for the difference equation plus any errors introduced by the numerical treatment of the boundary conditions. This is an error of approximation. 2. Round08 Error: The difference between the exact and machine solutions of the finite difference approximation. This is a consequence of the finite word length in any computer. This is an error of calculation. If we let E = exact (analytical) solution of the PDE = exact solution of the finite difference approximation N = numerical solution from a computer with finite accuracy

D then

Discretization error = E - D Roundoff error = E = N - D

(2.18)

Equation (2.18) can be written as (2.19)

N=D+E

where E is the roundoff error, which we will simply call "error" for brevity in the remainder of this section. The numerical solution N must satisfy the difference equation. Hence from Eq. (2.16), we have

D:"

+ E:+' At

- D: - E:

- 07+1+ & : + I

- 20:

- 24'

(W'

07-1

+ ~l-1 (2.20)

By definition, D is the exact solution of the difference equation, hence DY"

- D: - DY+l - 2 0 :

+ D;-1

(W2

(2.21)

Subtracting Eq. (2.21) from (2.20), we get (2.22)

Garg

Thus the error E also satisfies the difference equation. Let us now consider the stability of the difference Eq. (2.16). The solution of Eq. (2.16) will be stuble if the errors ci decay, or at best stay the same, as the solution progresses from time step n to n + 1. If the ci's grow larger during marching of the solution from step n to n + 1, the solution is unstuhle. Thus, for the solution to be stable, &;+I

-o. Cebeci and Smith (1974) developed a widely used version of the twolayer mixing-length model with (3.36) \pTo

= O.O168U,

b: FKleb

(3.37) (3.38)

Pressure gradient effects are incorporated through A+=

(3.39)

while intermittency effects at the boundary layer edge are described by FE;leb

[ + ($]-' 1

(3.40)

The eddy viscosity computation switches from the inner function to the outer function at the point where they are equal, which can be estimated (Wilcox, 1993a, p. 51) as =

0.04 Re,,

(3.41)

Baldwin and Lomax (1978) formulated another widely used two-layer eddy viscosity model that does not require an expensive calculation of the boundary layer thickness in order to determine a length scale in the outer region. Their model uses the distance away from the wall corresponding to the location of the maximum vorticity as the outer region length scale, making it especially attractive for use in a Navier-Stokes code.

Turbulence Modeling

(3.43) (3.44)

Fmax = max[y(l - e - y + I A + ) I cr) I ]

(3.45) (3.46)

where c,, = 1.6, CKleb = 0.3, Cwake= 1.0, ymaxis the location of FmaX, and is the maximum velocity. Granville (1987) has proposed modifications for these constants and factors to improve prediction of boundary layers under pressure gradients. Despite their obvious theoretical shortcomings, these algebraic models have been widely used due to their simplicity and robustness. Wilcox (1993a, pp. 53-64) presents a detailed comparison of results for channel and pipe flows and attached and separated boundary layers. Cebeci and Smith (1974) and Kays and Crawford (1993) describe a variety of modifications to incorporate blowing, suction, curvature, roughness, pressure gradients, and near-wall effects. Martinelli and Yakhot (1989) used an RNG algebraic model to predict the transonic flow over an airfoil with good results. Kirtley (1992) applied a similar model to the computation of the complex flow within a compressor rotor and showed improved results compared to the Baldwin-Lomax model. Despite having their constants derived analytically as part of the RNG procedure, both of these models are still dependent on the empirical specification of an appropriate length scale. -

U,,,

3.2.2 One-Equation Models The most obvious shortcoming of the algebraic eddy viscosity models discussed above is that they unrealistically compute vT = 0 and (?ii/i?jt = 0 at the centerline of a pipe or channel. Prandtl postulated the use of a turbulent velocity scale V T = JE to replace umix = lmiXi?ii/dyin (3.30) where k = (1/2)(u" + ZI'VI is the turbulent kinetic energy. The eddy viscosity is then computed as

+ m)

VT =

lJir

(3.47)

An exact governing equation for the transport of k can be derived by subtracting the Reynolds-averaged momentum equation from the instantaneous momentum equation and then Reynolds-averaging the result (Kays and

Schwab

Crawford, 1993, p. 56):

ak -+ Uj at

dk -= 8Xj

- aiii a -+; -- E + - [ aXj axj

ak v--axj

A &]

1U‘. U’.U’.2 P J

(3.48)

From left to right, the terms are unsteady rate, convection, production, dissipation, molecular diffusion, turbulent diffusion, and pressure diffusion. The last two terms are generally combined and modeled as a turbulent gradient-diffusion term using the eddy viscosity, while the dissipation is modeled as E = CDk3I2/1to yield (3.49)

The one-equation model offers some theoretical improvement in the computation of the eddy viscosity due to incorporating nonlocal and historical effects, such as the decay of freestream turbulence, through a transport equation. However, it still suffers from the requirement that the mixing length be specified, and thus offers no practical advantage over the algebraic models, especially in three-dimensional flows, where a plausible length scale can be quite difficult to define. Baldwin and Barth (1990) and Spalart and Allmaras (1992) have developed more complex versions of the one-equation model where the dependent variable is the eddy viscosity itself. Both models show some improvement in computing separate flows compared to algebraic models, but their performance on other flows remains mediocre.

3.2.3 Two-Equation Models Since the k-transport equation is well established as providing a reasonable turbulent velocity scale, the two-equation models offer the next step up in theoretical sophistication by supplying a transport equation for a second variable to establish the turbulent length scale. By an overwhelming margin, the most popular choice for the second variable is E, which appears directly in the k-transport equation; o = E/k and z = k/e are the bestknown alternatives. Jones and Launder (1972) developed what is generally considered to be the standard k-E model, although Chou (1945) and Harlow and Nakayama (1968) made earlier attempts at modeling the &-transportequation. An exact equation can be derived by differentiating the fluctuating velocity equation with respect to x i , multiplying the result by vaui/dxj, and finally Reynoldsaveraging the result (Kays and Crawford, 1993, pp. 59-60; Wilcox, 1993a, p. 88).

Turbulence Modeling

The exact equation contains many new correlation terms impossible to measure experimentally; however, some researchers (Mansour, Kim, and Moin, 1989; Rodi and Mansour, 1993) have attempted to directly model the terms in the etransport equation by using DNS results. The standard approach is to ignore the exact equation and to substitute a rational model equation, similar to the k-transport equation, that attempts to balance production, dissipation, convection, and diffusion to produce a reasonable profile of E that then produces reasonable profiles for k and v T . The standard model (Jones and Launder, 1972) consists of the following equations : VT

c, k2

(3.50) (3.51)

- + u j - = c,1 at

axj

where c, = 0.09, cC1= 1.55, cC2= 2.0, ok = 1.0, (T, = 1.3. Launder and Sharma (1974) revised two of the constants to their more generally accepted values: ct1 = 1.44 and cE2= 1.92. Wilcox (1993a, 1993b) has long been a proponent of the k - o model, k used as an alternative to the where a transport equation for o = ~ / is etransport equation. This model appears to offer improved performance in boundary layers under pressure gradients. (3.53) (3.54)

where a = 0.556, /?= 0.075, B* = 0.09, ok= 2.0, o, = 2.0. Even though both of the above models include viscous diffusion terms, their constants were evaluated for high-Reynolds-number flows, typically decaying hom*ogeneous turbulence and the logarithmic region of a boundary layer. Neither of the above models gives accurate solutions when integrated through the sublayer to the wall with no-slip boundary conditions.

Schwab

Any one of three following schemes is typically used to modify the models for the near-wall region. The first method uses the classic law of the wall to derive algebraic wall functions that specify values for k , E, and cu in terms of the friction velocity and the distance from the wall for the first grid point off the wall, presumably located in the logarithmic region. Obviously, this method is problematical for separated flows, and the location of the first grid point is largely empirical, even for attached boundary layers. Once the friction velocity is determined via iterative solution of the law of the wall at the first grid point, (3.56) the wall functions can be easily computed (Wilcox, 1993a, pp. 126-127) as (3.57) The second method uses a robust mixing-length model for the near-wall region. The two-equation model is matched to the mixing-length model at a point generally specified in terms of a fixed y f within the logarithmic region or a fixed value of v T / v large enough to ensure negligible viscous effects. The eddy viscosity computed from the mixing-length model is used to compute boundary values for k and E at the match point: (3.58) This method is more general than using wall functions, since it contains no explicit dependence on the law of the wall, which is only valid for attached two-dimensional shear layers. However, most mixing-length models use Van Driest damping through the sublayer and therefore become theoretically invalid for separated flows. The exact location of the match point also remains largely empirical. The most theoretically satisfactory scheme to include near-wall effects with two-equation models is to modify the standard model transport equations and allow integration right down to the wall with appropriate boundary conditions. This requires more grid points in the near-wall region than the other two near-wall schemes described above, but does permit a continuous solution without the problem of determining an artificial match point. Low-Reynolds-number (LRN) two-equation models are currently almost as ubiquitous as the algebraic models. Patel, Rodi, and Scheuerer

Turbulence Modeling

(1985) published a useful review of eight LRN two-equation models. Lang and Shih (1991), Thangam and Speziale (1992), Michelassi and Shih (1993), Steffen (1993), and Wilcox (1993a, pp. 138-146; 1993b) also published similar reviews. The general approach for the LRN models is to use damping functions to modify the major constants in the two-equation models to force the desired asymptotic behavior at the wall. These damping functions are typically designed as exponential functions of either y + = u,y/v, Re, = ,/kj./v, or ReT = k 2 / w to allow a smooth transition from the low-Reynolds-number behavior to the unmodified high-Reynolds-number behavior. Additional terms may also be added to the equations to adjust for a change in the dissipation variable to i = E - D,where D is a representation of the asymptotic value subtracted to allow a zero-valued boundary condition. Three of the most popular and well-tested LRN k-E models (Launder and Sharma, 1974; Lam and Bremhorst, 1981; Chien, 1982) are summarized in Table 1.

(3.60)

(3.61) For clarity, the functions are displayed in a form appropriate for a twodimensional boundary layer, where y is the direction normal to the flow. The Launder-Sharma model requires extra terms involving derivatives, including a second derivative of the velocity, since it exploits the zerovalued boundary condition for E. The Chien model uses nonderivative formulations for the additional terms. The Lam-Bremhorst uses the true boundary condition for E and thus requires no additional terms. While the LRN models above are reasonably robust, the damping functions mimic the desired near-wall behavior solely through viscous effects. The independent kinematic blocking effects of wall proximity on the eddy viscosity are not modeled. Myong and Kasagi (1990) developed a formulation forf, that accommodates the wall proximity effect through a rational blending of the dominant dissipation length scales near and away from the __ wall. Durbin (1991) proposed a novel method involving a third equation for dt’’ to provide a turbulent velocity scale that naturally incorporates both blocking and viscous effects near the wall.

Schwab

Table 3.1 LRN k - E Models Launder-Sharma

Chien

Lam-Bremhorst ~

0.09 1.44 1.92 1.o 1.3 exp( -3.4/[1 .O f,

D E E,,

+ Re,/50.0]2

1.o 1.0 - 0.3 exp( -Re+) 2v(aJi;/aY) * 2vv&%/ay)2 0

0.09 1.44 1.92 1.o 1.3 [1 .o - exp( -0.0165 Re,)]* x [1.0 20,5/Re,] 1.0 (0.05/r,)3 1 .O - exp( -Re+) 0 0 v ( a2k/ay 2,

0.09 1.35 1.80 1.o 1.3 1.O - exp( -0.01 15y+) 1 .o 1.O - 0.22 exp( - Re+/36.0) 2vk/y2 [ - 2vE/y2]exp( - 0.5y+) 0

Turbulence Modeling

Speziale and Thangam (1992) used an RNG k-E model to predict the separated flow over a backward-facing step with excellent results. The RNG model equations were identical to the standard model equations (3.50)(3.52), but the constants derived through the RNG procedure were c, = 0.085, cE1= 1.42 - [q(l - q/q0)3/[1 + /?q3], cE2-= 1.68, ck = 0.7179, and 6,= 0.7179, where q = Sk/q S = Sij = 0.5(i)Ui/i)xj+ C?Uj/2xi), qo = 4.38, and /? = 0.015. No damping functions were required to integrate the model equations directly to the wall.

J(2sijsij),

3.2.4 Compressibility Effects The algebraic models require no special modifications to adequately compute boundary layer flows without heat transfer up to Mach 5 and mixing layer flows up to Mach 1. Since the length scale used in oneequation models is essentially similar to the mixing length used in the algebraic models, one-equation models also require no modifications for the same types of compressible flows. Both classes of models incorporate compressibility effects solely through the variable mean density and ignore the effects of the fluctuating density, thus following Morkovin’s hypothesis (Morkovin, 1962) that the turbulence structure is unaffected by compressibility for boundary layers below Mach 5 or mixing layers below Mach 1, where the ratio of the density fluctuations to the mean density is small. For boundary layers above Mach 5 or mixing layers above Mach 1, Morkovin’s hypothesis of negligible density fluctuations becomes invalid. The presence of large pressure and temperature gradients due to shocks or significant heat transfer further complicates the turbulence modeling. Higher-order turbulence models are required for such complex compressible flows. Sarkar et al. (1989), Zeman (1990), and Wilcox (1992) proposed a useful modification to the k-E model that successfully accounts for compressibility effects. An additional dissipation modeling term is added to the k-equation by splitting the dissipation into two parts E = E, + & d , where E, is the solenoidal component unaffected by compressibility, and &d is the dilatational component directly dependent on the divergence of the fluctuating velocity, which vanishes for incompressible flows. The solenoidal components is computed from the usual &-transport equation, while the dilatational component is modeled as a function of the turbulence Mach number M , = &/U multiplying the solenoidal component. Sarkar et al. (1989) postulated the simple formulation (3.62)

Schwab

while Zeman (1990) proposed the more complex form

,/m)

with A = 0.60 and M,, = 0 . 2 5 d m ) for mixing layers and A = 0.66 and M,, = 0.25 for boundary layers. Wilcox (1992) later proposed ~d

1.5{M+- M + , } H ( M T - M T O ) E ,

(3.64)

with M , , = 0.25, which produced improved results for both mixing layers and boundary layers when used in the k-cu model.

3.3 STRESS TRANSPORT MODELS 3.3.1 Differential Stress Models

Lakshminarayana (1986), Launder (1989), and Hanjalic (1994) all make compelling arguments for using some form of Reynolds stress transport models to overcome the inherent limitations of using an isotropic eddy viscosity when computing complex flows. Only Reynolds stress transport models can directly capture the effects of additional strain rates, curvature, rotation, and stress anisotropy on the individual turbulent stress terms. However, the potential improvement in fidelity is offset by the increased complexity and computational cost of Reynolds stress transport models compared to algebraic or two-equation eddy viscosity models, since they require six equations (one for each turbulent stress) and a dissipation rate equation, which all incorporate complex modeling terms. The area of Reynolds stress transport modeling research is extremely active today, and most researchers remain optimistic that these models will eventually supplant the ubiquitous algebraic and two-equation eddy viscosity models as computational power continues to increase. Lumley (1983, 1992) has long argued for including formal mathematical constraints, such as realizability and material frame indifference. Speziale (1991, 1995) recently produced useful reviews concentrating on improved analytical development methods. So et al. (1991) reviewed near-wall Reynolds stress transport modeling. Demuren and Sarkar (1993a, 1993b) systematically evaluated various approaches to modeling the troublesome pressure-strain term. The exact form of the transport equation for R i j = pu:u; can be derived by Reynolds-averaging the product of the instantaneous momentum equa-

Turbulence Modeling

tion and the fluctuating velocity (Wilcox, 1993a, pp. 17-19):

From left to right, the terms represent unsteady rate, convection, production (two terms), viscous dissipation, pressure-strain redistribution, viscous diffusion, and turbulent diffusion (three terms). The pressure-strain and turbulent diffusion terms require modeling, while the viscous dissipation is obtained from its own transport equation similar to that developed for the k-E model. Launder, Reece, and Rodi (1975) developed a high-Reynolds-number form of a differential stress model (DSM) that has been adopted as a basic form by the majority of researchers in this area: aR.. aR.. 2 d + u , I = p i j - - pdij at axk 3

+ njj+ D;

(3.66) (3.67) (3.68)

p j j = - ( R j j E%+ R i j dXj

3) axi

(3.69)

where c1 = 1.5, c2 = 0.6, c, = 0.25, ce1 = 1.44, ce2 = 1.92, c, = 0.18. Many more complex forms for the pressure-strain term (3.67) and turbulent diffusion term (3.68) have been proposed to improve the DSM predictions (Jones and Musonge, 1988; Demuren and Sarkar, 1993a, 1993b; Speziale and Gatski, 1994; Speziale, 1995). Near-wall modifications to eliminate wall-function boundary conditions have been proposed by So and Yoo (1986), Shima, (1988), Launder and Shima (1989), h i and So (1990b), and Durbin (1993). Pollard and Martinuzzi (1989) compared the predictions of five differential stress models for pipe flow. Demuren (1990) computed flows in various channels with complex cross section. Shima (1993a, 1993b) applied the Launder and Shima (1989) DSM to boundary layers with

Schwab

periodic pressure gradient, transpiration, free-stream turbulence, streamwise curvature, and rotation to demonstrate its ability to capture complex physics. Lien and Leschziner (1994a, 1994b) presented an extensive description of their implementation and validation of a DSM. Rubinstein and Barton (1992) developed an RNG version of a DSM. Hanjalic (1994) showed selected cases where the DSM predictions of complex flows were superior to those from two-equation models. It is clear that there are currently many competing approaches to DSM development and that much future testing is required before a definitive assessment can be attempted. 3.3.2 Algebraic Stress Models

Since the DSM approach may remain prohibitively expensive for some time, it seems natural to attempt a simplified formulation that would improve upon the isotropic eddy viscosity models without the full cost of the additional partial differential equations required by the DSM. There are two distinct approaches to this problem. The first involves developing a nonlinear constitutive relation through formal expansions to extend the isotropic eddy viscosity produced by a two-equation model into an anisotropic form. The second involves simplification or truncation of the DSM equations into a system of coupled algebraic equations dependent on k , E , and R i j . Both approaches yield algebraic stress models (ASM) similar in performance and cost. Lumley (1970) and Pope (1975) were among the first to propose nonlinear extensions to the linear Boussinesq eddy viscosity relation. Yoshizawa (1984), Speziale (1987), Rubinstein and Barton (1990), Taulbee (1992), and Gatski and Speziale (1993) have further developed this theoretical expansion approach in various directions. Shih, Zhu, and Lumley (1994) have shown a variety of successful computations with their ASM model based on invariant theory and realizability constraints. This particular ASM is described below as an example; it is similar to those of Yoshizawa (1984) and Rubinstein and Barton (1990) in its quadratic tensorial form.

aii.

2 aii, - - - 6ij ax, 3 ax,

aii.

2+ A

axj

(3.71)

Turbulence Modeling

where 2a,

Cr1

A, k q = s -,

+ q 3 + C3'

2a, =

Cr2

+ q3 + C3'

1 s = JG., sij = 2

aii, axj

2a,

Cr 3

+ q3 + C3

(3.72)

The first two terms in (3.71) are identical to the standard isotropic eddy viscosity, except that c, is now a function of the time scale ratio q. The constants in the k-transport and &-transport equations take on standard values, while the new constants are calibrated against experimental data. While this type of ASM formulation has an interesting theoretical basis, it will require more testing for a definitive assessment of practical utility. Rodi (1976) developed a popular simplification of the DSM by assuming that the net convective and diffusive transport ofufuj is locally proportional to the net convective and diffusive transport of k, thus reducing the set of coupled partial differential equations to a set of coupled algebraic equations. Rodi proposed that the local coeficient of proportionleading to the following set of equations for ality is the ratio

mj/k, a -;;- &] k[ -

u:,u:,

mj:

(3.74)

Appropriate models for E ~ , and l l i j can now be inserted, along with local values of k and E computed from their transport equations. Launder (1982) proposed a generalized version of this type of ASM to allow preferential transport of the shear stresses compared to the normal stresses. Lakshminarayana (1986) and Launder (1989) present extensive reviews of this type of ASM, which can also be formulated to include curvature and rotation effects. The ASM formulation is appropriate only for high-

Schwab

Reynolds-number flows; thus, it must be matched to a low-Reynoldsnumber model for the near-wall region. It should also be noted that the ASM is notoriously stiff, often requiring extreme underrelaxation. 3.3.3

Compressibility Effects

Few researchers have looked into compressibility effects for DSM. Zhang et al. (1993) showed good results for supersonic boundary layers with adiabatic and cooled walls using a near-wall DSM with no compressibility modifications other than including the variable mean density. The general approach would be to obviously include the direct effects of variable mean density along with the extra dilatational dissipation (Sarkar et al., 1989; Zeman, 1990), as done for the two-equation models. It seems reasonable to presume that new forms of the pressure-strain model will also be required, since most current formulations depend upon the classic Chou (1945) technique involving an incompressible Poisson equation for the pressure. 3.4

THERMAL MODELS

The recent literature shows relatively few descriptions of thermal turbulence modeling, compared to the enormous number of works on dynamic turbulence modeling. Benocci (1991) published a review that concentrated on dynamic turbulence modeling for flows with heat transfer. Launder (1993) argued that second-moment models for the Reynolds stresses should remain the most important part of computing turbulent heat transfer, since the turbulent velocity field appears to have much more influence than the turbulent temperature field. So, Yuan, and Sommer (1992) developed a similar conclusion in a presentation of their hierarchy of thermal turbulence models, although they pointed to the separate computation of thermal turbulence as their goal for ultimate fidelity. Since the computation of the turbulent temperature field requires a suitable computation of the turbulent velocity field, it would be illogical to use a higher-order model for the temperature field with a lower-order model for the velocity field. However, it seems extremely optimistic to rely on lower-level models for the temperature field with higher-order models for the velocity field in complex flows, where the velocity and temperature fields may involve entirely different physical mechanisms. Although it remains relatively inferior to dynamic turbulence modeling, the continued development of thermal turbulence modeling can offer conceptually parallel modeling of the turbulent velocity and temperature fields, thus permitting more generalized and robust computation of complex flows with heat transfer .

Turbulence Modeling

Figure 3.1

Turbulent Prandtl number for a boundary layer at Re, = 670

This section will examine a hierarchy of approaches to thermal turbulence modeling, with particular emphasis on their potential to capture physical effects. It will be limited to fully turbulent, non reacting, singlephase, nonbuoyant flows. The effects of molecular Prandtl number will not be explicitly considered, since they are relatively unimportant, except for liquid metals, where molecular conduction can dominate turbulent diffusion throughout the entire flow field. Although they are extremely interesting and important, transitional flows will not be considered in this Chapter, since their complexity requires a separate detailed treatment beyond the scope of the present work. For the current work, a fundamental distinction will be utilized to classify thermal turbulence models as either specification models or transport models. Specification models involve the computation of the turbulent thermal diffusivity through the specification of a turbulent Prandtl number, which may be an empirical constant or a complex anisotropic function. Transport models involve the actual computation of thermal turbulence variables through partial differential equations (PDE) governing their transport.

Schwa b 2.0

0 I .5

Kasajy DNS

- - - - - Yakhot-Orszag

1.0

0.5

0.0

100

150

Y+ Figure 3.2 Turbulent Prandtl number for a channel flow at Re, = 150.

3.4.1 Specification Models Constant Pr,

The concept of a constant turbulent Prandtl number is too valuable to abandon for many engineering computations, especially those involving simple boundary layers or fully developed duct flows. Kays and Crawford (1993) and Zukauskas, Slanciauskas, and Karni (1987) show a variety of such computations exhibiting good agreement with experimental data. Launder (1978) cites ubiquitous values of 0.9 for wall-bounded flows and 0.7 for free shear flows. In a variety of experimental studies summarized by Kays (1994), Pr, appears to remain fairly constant through the log region of zero-pressuregradient (ZPG) boundary layers. In fact, Pr, can be directly computed from the slopes of the nondimensionalized velocity and temperature profiles if one makes the customary assumption of constant turbulent shear stress and heat flux in this region. The evidence supporting a constant Pr, begins to diminish near the wall or in flows with pressure gradients. We find conflicting results as we examine Pr, near the wall. Although many experiments, such as those of Antonia and Kim (1991) and Bagheri,

Turbulence Modeling

Strataridakis, and White (1992), show a definite trend of increased Pr, within the sublayer, we must remember that this is an extremely difficult measurement with relatively large experimental uncertainties. Recent direct numerical simulation (DNS) results presented in Figs. 3.1 and 3.2, for a ZPG boundary layer by Bell and Ferziger (1993) and for a channel flow by Kasagi, Tomita, and Kuroda (1992), show that Pr, tends toward a constant value near unity through the sublayer. We note that the DNS results were computed at relatively low Reynolds numbers, where the sublayer is relatively thick, which may influence Pr, in this region. Figures 3.3 and 3.4 show a clear trend of reduced Pr, across an adverse-pressure-gradient (APG) boundary layer (Blackwell, Kays, and Moffat, 1972) and enhanced Pr, across a favorable-pressure-gradient (FPG) boundary layer (Roganov et al., 1984), compared to their respective ZPG cases. Despite this evidence of Pr, variation with applied pressure gradient, Huang and Bradshaw (1995) show good agreement with experimental temperature profiles using a constant Pr,. The crucial aspect appears to be the

Blackwell ZPG

Blackwell APG

0.0 0

100

200

300

400

500

600

Y+ Figure 3.3 layer.

Turbulent Prandtl number for an adverse-pressure-gradient boundary

100

Schwab

0 0

0.5

RoganovFPG

O 0 o

1 0

RoganovZPG

100

200

400

300

500

600

700

Y+ Figure 3.4 layer.

Turbulent Prandtl number for a favorable-pressure-gradient boundary

ability of the dynamic turbulence model to model the increased sublayer thickness under F P G and the decreased sublayer thickness under APG.

Variable Pr,

Since the assumption of a constant Pr, permits generally accurate predictions of the temperature profiles in simple shear layers, but fails to match the observed variation in Pr, through the shear layer, we next examine the approach of modeling Pr, as a variable. Reynolds (1975) published the classic review of over 30 variable Pr, models over 20 years ago. Two of the most popular modern approaches are discussed below. The explicit empirical formula of Kays and Crawford (1993, pp. 266268) depends upon the turbulent Peclet number, an empirical freestream turbulent Prandtl number, and an empirical constant.

101

Turbulence Modeling

+ c Pe,/&

Pe, = - Pr, V

PrTm=

- (C

Ped2 (3.75)

= 0.86

(air),

= 0.2

(air)

For Pr near unity, it predicts Pr, to be nearly constant through the log region, but rapidly increasing toward the wall through the sublayer, thus matching the general trend of the available experimental data, but not the DNS results shown in Figs. 3.1 and 3.2. The implicit model developed by Yakhot and Orszag (1986) is based on renormalization group theory (RNG) and depends on the turbulent Peclet number and molecular Prandtl number; the constants that appear in the model are derived from fundamental analysis, not from empiricism.

{

l/PrefL- 1.18)00’{ 1 1/Pr 1.18 1/Pr + 2.18

Preff

Pr + Pe, 1 + Pe,/Pr,’

1 1 + Pe,/Pr

(3.76)

VT Pe, = Pr

For Pr near unity, it predicts Pr, to remain nearly constant throughout the entire shear layer, as shown in Figs. 3.1 and 3.2 Since both the Kays-Crawford and Yakhot-Orszag models predict equally good results for the temperature profile in simple shear layers for air using entirely different approaches, we conclude that specification of a variable Pr, through the shear layer is of limited utility when Pr is near unity. The variable Pr, models should be recognized primarily for their ability to predict the higher values of Pr, associated with low Pr liquid metals. Generalized Gradient Diffusion Hypothesis (GGDH)

In order to address the obvious shortcomings of specifying an isotropic Pr, when an anisotropic dynamic turbulence model is used, we now examine a model first described by Daly and Harlow (1970) and later extensively used by Launder (1988) and others. The generalized form of the gradient diffusion hypothesis yields (3.77)

102

Schwab

(3.78) where the ratio k/u:u) can be regarded as an anisotropic turbulent Prandtl number relating an anisotropic thermal diffusivity to an isotropic momentum diffusivity. If the anisotropic turbulent stresses are accurately predicted, this form offers some potential for capturing anisotropy in the turbulent fluxes. Since the turbulent fluxes remain directly dependent on the turbulent stresses, we can view the GGDH as a special form of a specification model. Launder (1993, 1988) has used this model extensively with a differential stress transport model to reasonably capture the anisotropic effects of rotation and curvature as they are manifested in the turbulent stresses. 3.4.2

Eddy Diffusivity Models

Turbulent Diffusivity Concept without Pr,

We now turn our attention to models involving the computation of actual thermal turbulence variables. Both the one-equation and two-equation thermal turbulence models involve computing a turbulent thermal diffusivity a, without requiring specification of Pr,. We assume that a,, analogous to v T , is proportional to some fundamental turbulent scales: a, K [velocity] x [length] cc [velocity12 x [time]

(3.79)

The obvious choice for the square of the velocity scale is the turbulent kinetic energy: [velocity]’ cc k

(3.80)

It then seems reasonable to propose that the time scale reflects some combination of the time scales for turbulent momentum transport and turbulent thermal transport: [time] cc zM$

(:)‘(2>”.

(3.81)

We can now formulate the thermal diffusivity in terms of the momentum diffusivity and the time scale ratio R = z,/z: (3.82) Commonly proposed values for N include 0.5 (Nagano and Kim, 1988; Sommer, So, and Zhang, 1993; Hattori, Nagano, and Tagawa, 1993), 2.0

Turbulence Model i ng

103

(Youssef, Nagano, and Tagawa, 1992; Yoshizawa, 1988), and 1.0 (Schwab and Lakshminarayana, 1994; Schwab and Lakshminarayana, 1995). As R departs from its equilibrium value of 0.5 (Beguier, Dekeyser, and Launder, 1978), we can see that the choice of N will produce clearly different behavior for aT. One-Equation Models

The simplest thermal transport model involves only a single partial differential equation governing the transport of the fluctuating temperature variance k, = 8‘8’/2: (3.83)

We can substitute the following relation : (3.84)

but we still have to estimate a value for the time scale ratio, R . Although this model is attractive in its simplicity, and allows separate evolution of the thermal turbulence, the required specification of R remains a severe limiting factor. Two-Equation Models

The obvious next level of fidelity in modeling involves the computation of C e using a separate transport PDE. While an exact PDE for C e can be derived, it governs the small-scale dissipation process, rather than the large-scale energy transfer process that is more congruent with other modeling assumptions. Therefore, a model PDE is generally used to attempt to empirically balance the convection, production, dissipation, and diffusion terms; the original formulation by Newman, Launder, and Lumley (1981) was later modified by Nagano and Kim (1988) to include the second production term :

- c4f4 k +

z)$1

(3.85) ke [(a + We note that the PDE contains multiple production and dissipation terms corresponding to both dynamic and thermal processes. - c3f3

Schwab

104

Schwab and Lakshminarayana (1994, 1995) developed an alternative two-equation model by constructing a PDE for = ke/Ee as the second equation in order to exploit the zero-valued wall boundary condition and simple monotonic behavior.

+axi [ ( a

2)21

The direct gradient transport model used for the first diffusion term requires the two additional cross-diffusion terms for consistency. Schwab and Lakshminarayana (1995) compared the predictions of four low-Reynolds-number two-equation thermal models for two-dimensional channel flows with heat transfer. The results predicted by the two-equation thermal models show generally improved predictions compared to the constant Pr, model. Further predictions of more complex flows with heat transfer will be required to show the improved fidelity possible compared to the constant Pr, model. Since the near-wall flow structure is extremely important for the computation of surface heat transfer, the near-wall turbulence modeling can be crucial. Although thermal turbulence wall functions can be developed in a manner similar to those developed for the k-E model, this approach seems disjointed. If a separate thermal turbulence transport model is used to provide improved thermal predictions, it should incorporate low-Reynoldsnumber effects rather than relying upon wall functions.

3.4.3 Flux Transport Models

Flux transport models, with an individual PDE governing the transport of each turbulent flux, represent the ultimate fidelity in thermal turbulence from a concurrent models. They require the accurate prediction of turbulent stress transport model, so they have the highest potential for capturing anisotropic interactions between the turbulent fluxes and stresses that we might expect in complex flows subjected to rotation and curvature. Unfortunately, the development of flux transport models greatly lags the development of stress transport models, especially for predictions in the near-wall region. The models of Shih and Lumley (1986), Jones and Musonge (1988), and Lai and So (1990a) have achieved reasonable success,

Turbulence Modeling

105

but much more development and validation will be required to bring them to a level suitable for practical applications. The exact transport equation can be written as

or more compactly as (3.88) where the right-hand-side terms represent production by mean temperature gradients, production by mean velocity gradients, molecular dissipation, pressure scrambling, molecular diffusion, and turbulent diffusion. The modeling of the pressure scrambling and molecular dissipation terms is considered to be especially critical in the near-wall region (Lai and So, 1990a). So, Yuan, and Sommer (1992) found that the prediction of their combined stress transport and flux transport model was inferior to that of their combined two-equation models when compared to DNS results for a channel flow. They attribute this situation to the more extensive validation of the two-equation models compared to the stress and flux transport models, which have only recently been adapted for low-Reynolds-number near-wall flow regions. If we assume that we can ignore the viscous diffusion and molecular dissipation terms for high-Reynolds-number flow, we can obtain an algebraic form of the transport equation by postulating that the combined convection and turbulent diffusion terms for @ are proportional to those for the product fie, and that the proportionality factor is q/p0 (Gibson, 1978; Launder, 1988):

(3.89) After substituting back into the@ model f o r m :

PDE, we obtain an entirely algebraic

Schwab

106

This model requires the computation of k , e, k , , and E @ , and is limited to high-Reynolds-number flows by the underlying assumption. It would seem reasonable to combine this algebraic flux model with an algebraic stress model and appropriate low-Reynolds-number two-equation dynamic and thermal models. However, algebraic stress and flux models are notoriously stiff, often requiring extreme under relaxation that may increase the computational effort almost as much as the additional effort that differential flux transport models would require. The current research trend appears to be either to step down to a GGDH model or step up to a differential flux transport model, rather than continuing to develop an algebraic flux model. 3.4.4 Compressibility Effects

Since few researchers have even considered thermal turbulence transport models, it is to be expected that even fewer have extended them to compressible flows. At the current state of development and validation for thermal turbulence models, the incorporation of gross compressibility effects through the variable mean density appears to be a reasonable approach. Sommer, So, and Zhang (1993) have shown some improved results for high-Mach-number boundary layer flows over adiabatic and cooled walls using their compressible two-equation thermal turbulence model compared to a constant turbulent Prandtl number. 3.5

CONCLUSIONS AND RECOMMENDATIONS

The algebraic eddy viscosity models are computationally robust and economical. They are generally adequate for two-dimensional shear flows with mild pressure gradients. For more complex flows involving multiple strain rates, the mixing-length concept derived for a simple shear layer becomes inadequate. The one-equation models offer some theoretical improvement in the computation of the eddy viscosity due to incorporating nonlocal and historical effects, such as the decay of freestream turbulence, through a transport equation. However, they still suffer from the requirement that the mixing length be specified, and thus offer no practical advantage over the algebraic models, especially in three-dimensional flows, where a plausible length scale can be quite difficult to define. The two-equation models are nearly as ubiquitous as the algebraic models for practical engineering computations. Extensive testing has been reported, especially for the low-Reynolds-number versions that eliminate wall functions. They are adequate for two-dimensional flows with pressure

Turbulence Modeli ng

107

gradients and two-dimensional recirculating flows. While offering some distinct advantages over the algebraic and one-equation models, the twoequation models become inadequate whenever the underlying assumption of an isotropic eddy viscosity fails in highly three-dimensional flows with multiple strain rates. Coupling a two-equation model to an algebraic stress model offers the potential to capture some anisotropic effects, especially rotation and curvature. The differential stress models are the only modeling approach capable of directly capturing the anisotropic effects of multiple strain rates upon each individual turbulent stress. Very complex, highly three-dimensional flows may require this level of modeling for accurate predictions, but it remains computationally expensive and will not soon replace lower levels of modeling for most practical engineering computations. Many researchers are currently investigating the difficult pressure-strain term and near-wall effects; thus, extensive testing will be required for any definitive assessment. The algebraic stress models are an economical alternative to the differential stress models. Either the nonlinear constitutive relation approach or the proportional transport approach can be formulated to include rotation and curvature effects, thus extending the isotropic eddy viscosity limitations of the two-equation models to allow capturing some anisotropic effects. This level of modeling can provide good engineering predictions when isotropic two-equation models are obviously inadequate, and is recommended until the differential stress models become practical for routine computations. The specification of a constant Pr, remains extremely useful with isotropic eddy viscosity models, and is quite adequate for boundary layers and fully developed duct flows without anisotropic curvature or rotation effects. Pressure gradient effects are not explicitly captured with a constant Pr,, but correct temperature profiles can be predicted if the effective momentum sublayer thickness is made a function of the pressure gradient. The specification of a variable Pr, as a function of the turbulent Peclet number appears to have little discernible effect, since the proposed functions tend toward a constant value away from the wall for Prandtl numbers near unity. Such models should be primarily recognized for their ability to predict the higher values of Pr, associated with low-Prandtl-number liquid metal flows. The generalized gradient diffusion hypothesis offers a rational relation of turbulent fluxes to turbulent stresses and can incorporate anisotropic interactions with turbulent stresses affected by rotation and curvature. It is recommended for use with stress transport models until flux transport models reach sufficient maturity. The one-equation ke transport model eliminates the specification of Pr,

Schwab

108

and allows separate evolution of the thermal turbulence. However, its utility is severely limited by the required specification of the dynamic/thermal time scale ratio in order to relate the thermal dissipation to the momentum dissipation. Since we know that the dissipation plays a crucial role in turbulent transport, and we are attempting to model the thermal transport separately from the momentum transport, we should therefore model the thermal dissipation separately from the momentum dissipation. The two-equation transport models eliminate the specification of Pr,. and allow separate evolution of the thermal turbulence in a complete manner conceptually parallel to the well-tested two-equation dynamic models. They can be used through the near-wall region and offer the potential to incorporate anisotropic effects when used with algebraic stress and flux transport models. The two-equation level of thermal turbulence modeling is recommended for further exploration and application in complex flows where the physical processes affecting momentum transport and thermal transport may be quite different. The flux transport models represent the ultimate fidelity in modeling thermal turbulence. They require accurate prediction of the turbulent stresses through a concurrent stress transport model, and have the potential to directly capture the anisotropic interaction effects caused by rotation and curvature. However, the development and validation of flux transport models lag far behind that of stress transport models, especially for incorporating low-Reynolds-number effects in the near-wall region. The algebraic flux model offers a simpler formulation for use with a two-equation thermal model. Until the flux transport PDE models reach sufficient maturity, the generalized gradient diffusion hypothesis model is recommended for use with stress transport models. NOMENCLATURE

local speed of sound van Driest constant turbulent kinetic energy; thermal conductivity turbulent temperature variance length scale turbulent Mach number: M, = @/a pressure turbulent Peclet number: Pe, = Pr v,/v Prandtl number: Pr = v/o! turbulent Prandtl number: Pr, = vT = a, instantaneous value of flow variable

109

Turbulence Modeling

mean component of flow variable (Reynolds averaging) mean component of flow variable (Favre averaging) fluctuating component of flow variable (Reynolds averaging) fluctuating component of flow variable (Favre averaging) heat flux vector turbulent time scale ratio: R = z,/z Reynolds stress tensor turbulent Reynolds number: Re, = &y/v turbulent Reynolds number: Re, = k2/v& displacement thickness Reynolds number: Re,, = u,S*/v time Cartesian velocity components p friction velocity: U, = a transpiration velocity Cartesian space coordinates sublayer-scaled wall-normal distance: y = U, y/v +

thermal diffusivity turbulent thermal diffusivity specific heat ratio boundary layer thickness Kronecker delta function boundary layer displacement thickness boundary layer velocity thickness dissipation rate of turbulent kinetic energy permutation tensor dissipation rate of turbulent temperature variance temperature von Karman constant Escudier constant dynamic viscosity kinematic viscosity: v = p / p turbulent viscosity density turbulent time scale: z = k/E turbulent thermal time scale: z, = k , / ~ , viscous stress tensor wall shear stress dissipation function vorticity; specific dissipation rate: w = ~ / k

110

Schwab

REFERENCES Anderson DA, Tannehill JC, Pletcher RH. Computational Fluid Mechanics and Heat Transfer. Hemisphere, 1984. Antonia RA, Kim J. Turbulent Prandtl number in the near-wall region of a turbulent channel flow. Int J Heat Mass Transfer 34: 1905-1908,1991. Bagheri N, Strataridakis CJ, White BR. Measurements of turbulent boundary layer Prandtl numbers and space-time temperature correlations. AIAA J 30: 35-42, 1992. Baldwin BS, Barth TJ. A one-equation turbulence transport model for high Reynolds number wall-bounded flows. NASA TM- 102847, 1990. Baldwin BS, Lomax H. Thin layer approximation and algebraic model for separated turbulent flows. AIAA 78-257, 1978. Beguier C, Dekeyser I, Launder BE. Ratio of scalar and velocity dissipation time scales in shear flow turbulence. Phys Fluids 21 : 307-310, 1978. Bell DM, Ferziger JH. Turbulent boundary layer DNS with passive scalars. In: So RMC, Speziale CG, Launder BE, eds. Near-Wall Turbulent Flows. Elsevier, 1993, pp 327-336. Benocci C, Modeling of turbulent heat transport. VKI Technical Memorandum 47, 1991. Blackwell BF, Kays WM, Moffat RJ. The turbulent boundary layer on a porous plate: an experimental study of the heat transfer behavior with adverse pressure gradients. NASA CR-130291, 1972. Cebeci T, Smith AMO. Analysis of Turbulent Boundary Layers, Applied Mathematics and Mechanics. Vol. 15. Academic Press, 1974. Chien K-Y. Predictions of channel and boundary-layer flows with a low Reynoldsnumber turbulence model. AIAA J 20: 33-38, 1982. Chou PY. On the velocity correlations and the solution of the equations of turbulent fluctuation. Quart J Appl Math 3: 38-54, 1945. Daly BJ, Harlow FH. Transport equations in turbulence. Phys Fluids 13: 26342649, 1970. Demuren AO. Calculation of turbulent flow in complex geometries with a secondmoment closure model. ASME FED-Vol. 94. In: Bower WM, Morris MJ, Saming M, eds. Forum on Turbulent Flows-1990. 1990, pp 163-167. Demuren AO, Sarkar S. Perspective: systematic study of Reynolds stress closure models in the computations of plane channel flows. ASME J Fluids Eng 115: 5-12 1993a. Demuren AO, Sarkar S. Study of second-moment closure models in computations of turbulent shear flows. In: Rodi W, Martelli F, eds. Engineering Turbulence Modeling and Experiments 2. Elsevier, pp 1993b, 53-62. Durbin PA. Near-wall turbulence closure modeling without damping functions. Theoret Comput Fluid Dynamics 3: 1-13, 1991. Durbin PA. A Reynolds stress model for near-wall turbulence. J Fluid Mech 249: 465-498, 1993.

Turbulence Modeling

111

Gatski TB, Speziale CG. On explicit algebraic stress models for complex turbulent flows. J Fluid Mech 254: 59-78, 1993. German0 M, Piomelli U, Moin P, Cabot WH. A dynamic subgrid-scale eddy viscosity model. Phys Fluids A 3: 1760-1765, 1991. Gibson MM. An algebraic stress and heat-flux model for turbulent shear flow with streamline curvature. Int J Heat Mass Transfer 21 : 1609-1617, 1978. Granville PS. Baldwin-Lomax factors for turbulent boundary layers in pressure gradients. AIAA J 25: 1624-1627,1987. Hanjalic K. Advanced turbulence closure models: a view of current status and future prospects. Int J Heat Fluid Flow 15: 178-203, 1994. Harlow FH, Nakayama PI. Transport of turbulence energy decays rate. University of California Report LA-3854, 1968. Hattori H, Nagano Y, Tagawa M. Analysis of turbulent heat transfer under various thermal conditions with two-equation models. In: Rodi W, Martelli F, eds. Engineering Turbulence Modeling and Experiments 2. Elsevier, 1993, pp 43-52. Huang PG, Bradshaw P. Law of the wall for turbulent flows in pressure gradients. AIAA J 33: 624-632,1995. Jones WP, Launder BE. The prediction of laminarization with a two-equation model of turbulence. Int J Heat Mass Transfer 15: 301-314, 1972. Jones WP, Musonge P. Closure of the Reynolds stress and scalar flux equations. Phys Fluids 31 : 3589-3604,1988. Kasagi N, Tomita Y, Kuroda A. Direct numerical simulation of passive scalar field in a turbulent channel flow. ASME J Heat Transfer 114: 598-606, 1992. Kays WM. Turbulent Prandtl number-where are we?, ASME J Heat Transfer 116: 284-295,1994. Kays WM, Crawford ME. Convective Heat and Mass Transfer. 3rd ed. McGrawHill, 1993. Kays WM, Moffat RJ. The behavior of transpired turbulent boundary layers. In: Studies in Convection: Theory, Measurement, and Applications, Vol. 1 . Academic Press, 1975, pp 213-319. Kirtley KR. An algebraic RNG-based turbulence model for three-dimensional turbomachinery flows. AIAA J 30: 1500-1506,1992. Lai YG, So RMC. Near-wall modeling of turbulent heat fluxes. Int J Heat Mass Transfer 33: 1429-1440, 1990a. Lai YG, So RMC. On near-wall turbulent flow modeling, J Fluid Mech 221 : 641673, 1990b. Lakshminarayana B. Turbulence modeling for complex shear flows. AIAA J 24: 1900-1917, 1986. Lam CHG, Bremhorst KA. Modified form of the k-E model for predicting wall turbulence. ASME J Fluids Eng 103: 456-460, 1981. Lang NJ, Shih T-H. A critical comparison of two-equation turbulence models. NASA TM-105237, 1991. Launder BE. Heat and mass transport. In: Bradshaw P, ed. Turbulence. 2nd ed.

112

Schwab

Topics in Applied Physics. Vol. 12. Springer-Verlag, 1978, Chap. 6, pp 231 287. Launder BE. A generalized algebraic stress transport hypothesis. AIAA J 20: 436 437, 1982. Launder BE. On the computation of convective heat transfer in complex turbulent flows. ASME J Heat Transfer 110: 1 1 12-1 128, 1988. Launder BE. Second-moment closure and its use in modeling turbulent industrial flows. Int J Numer Meth Fluids. 9: 963-985, 1989. Launder BE. Modeling convective heat transfer in complex turbulent flows. In: Rodi W, Martelli F, eds. Engineering Turbulence Modeling and Experiments 2. Elsevier, 1993, pp 3-22. Launder BE, Sharma BI. Application of the energy dissipation model of turbulence to the calculation of flow near a spinning disk. Lett Heat Mass Transfer 1 : 131-138, 1974. Launder BE, Shima N. Second-moment closure for the near-wall sublayer: development and application. AIAA J 27 : 1319-1 325, 1989. Launder BE, Reece GJ, Rodi W. Progress in the development of a Reynolds stress turbulence model. J Fluid Mech 68: 537-566, 1975. Lien FS, Leschziner MA. A general non-orthogonal collocated finite volume algorithm for turbulent flow at all speeds incorporating second-moment turbulence-transport closure. Part 1: Computational implementation. Comput Meth Appl Mech Eng 114: 123-148, 1994a. Lien FS, Leschziner MA. A general non-orthogonal collocated finite volume algorithm for turbulent flow at all speeds incorporating second-moment turbulence-transport closure. Part 2: Application. Comput Meth Appl Mech Eng 114: 149-167, 1994b. Lumley JL. Toward a turbulent constitutive relation. J Fluid Mech 41: 413-434, 1970. Lumley JL. Turbulence modeling. ASME J Appl Mech 50: 1097-1 103, 1983. Lumley JL. Some comments on turbulence. Phys Fluids A 4:2: 203-21 1, 1992. Mansour NN, Kim J, Moin P. Near-wall k-E turbulence modeling. AIAA J 27: 1068-1073, 1989. Martinelli L, Yakhot V. RNG-based turbulence transport approximations with applications to transonic flows. AIAA 89- 1950-CP, 1989. Michelassi V, Shih T-H. Elliptic flow computation by low Reynolds number twoequation turbulence models. NASA TM- 105376, 1993. Morkovin MV. Effects of compressibility on turbulent flow. In: Favre A, ed. The Mechanics of Turbulence. Gordon and Breach, 1962, pp 367--380. Myong HK, Kasagi N. A new approach to the improvement of k-E turbulence model for wall-bounded shear flows. AMSE Int J 33: 63--72, 1990. Nagano Y, Kim C. A two-equation model for heat transport in wall turbulent shear flows. ASME J Heat Transfer 110: 583-589, 1988. Newman GR, Launder BE, Lumley JL. Modeling the behaviour of hom*ogeneous scalar turbulence. J Fluid Mech 1 1 1: 2 17-232, 1981. Pate1 VC, Rodi W, Scheuerer G. Turbulence models for near-wall and l o w Reynolds number flows: a review. AIAA J 23: 1308-1319, 1985.

Turbulence Modeling

113

Pollard A, Martinuzzi R. Comparative study of turbulence models in predicting turbulent pipe flow. Part 11: Reynolds stress and k-E models. AIAA J 27: 1714-1 721, 1989. Pope SB. A more general effective-viscosity hypothesis. J Fluid Mech 72: 331-340, 1975. Rai MM, Moin P. Direct simulations of turbulent flow using finite-difference schemes. J Comput Phys 96: 15-53, 1991. Reynolds AJ. The prediction of turbulent Prandtl and Schmidt numbers. Int J Heat Mass Transfer 18: 1055-1069, 1975. Rodi W. A new algebraic relation for calculating the Reynolds stresses. ZAMM 56: T219-T221, 1976. Rodi W, Mansour NN. Low Reynolds number k-E modeling with the aid of direct simulation data. J Fluid Mech 250: 509-524, 1993. Rogallo RS, Moin P. Numerical simulation of turbulent flows. Annu Rev Fluid Mech 16: 99-137, 1984. Roganov PS, Zabolotsky VP, Shishov EV, Leontiev AI. Some aspects of turbulent heat transfer in accelerated flows on permeable surfaces. Int J Heat Mass Transfer 27: 1251-1259, 1984. Rubinstein R, Barton JM. Nonlinear Reynolds stress models and the renormalization group. Phys Fluids A 2: 1472-1476, 1990. Rubinstein R, Barton JM. Renormalization group analysis of the Reynolds stress transport equation. NASA TM-105588, 1992. Sarkar S, Erlebacher G, Hussaini MY, Kreiss HO. The analysis and modeling of dilatational terms in compressible hom*ogeneous turbulence. NASA CR181959,1989. Schwab JR, Lakshminarayana B. Dynamic and thermal turbulent time scale modeling for hom*ogeneous shear flows. (NASA TM-106635), ASME FED-Vol. 184. In: Donovan JF, Dutton JC, eds Boundary Layer and Free Shear Flows. 1994, pp 75-86. Schwab JR, Lakshminarayana B. Dynamic and thermal turbulent time scale modeling for wall-bounded shear flows. ASME HTD-Vol. 318. In: Anand NK, Amano RS, Armaly BF, eds. Heat Transfer in Turbulent Flows. 1995, pp 111-1 18. Shih T-H, Lumley JL. Influence of timescale ratio on scalar flux relaxation: modeling Sirivat & Warhaft’s hom*ogeneous passive scalar fluctuations. J Fluid Mech 162: 21 1-222, 1986. Shih T-H, Zhu J, Lumley JL. Modeling of wall-bounded complex flows and free shear flows (NASA TM-106513), ASME FED-Vol. 184. In: Donovan JF, Dutton JC, eds. Boundary Layer and Free Shear Flows. 1994, pp 105-1 12. Shima N. A Reynolds-stress model for near-wall and low-Reynolds-number regions. ASME J Fluids Eng 110: 38-44, 1988. Shima N. Prediction of turbulent boundary layers with a second-moment closure. Part 1. Effects of periodic pressure gradient, wall transpiration, and freestream turbulence. ASME J Fluids Eng 115: pp. 56-63, 1993a.

114

Schwab

Shima N. Prediction of turbulent boundary layers with a second-moment closure. Part 2. Effects of streamwise curvature and spanwise rotation. ASME J Fluids Eng 115: 64-69, 1993b. So RMC, Yoo GJ. On the modeling of low-Reynolds-number regions. NASA CR3994,1986. So RMC, Yuan SP, Sommer TP. A hierarchy of near-wall closures for turbulent heat transfer. Trends Heat, Mass, Momentum Transfer 2: 203-221, 1992. So RMC, Lai YG, Zhang HS, Hwang BC. Second-order near-wall turbulence closures: a review. AIAA J 29: 1819-1835, 1991. Sommer TP, So RMC, Zhang HS. Near-wall variable-Prandtl-number turbulence model for compressible flows. AIAA J 31 : 27-35, 1993. Spalart PR, Allmaras SR. A one-equation turbulence model for aerodynamic flows. AIAA 92-439, 1992. Speziale CG, On nonlinear k-I and k-E models of turbulence. J Fluid Mech 178: 459-475, 1987. Speziale CG. Analytical methods for the development of Reynolds-stress closures in turbulence. Ann Rev Fluid Mech 23: 107-157. Speziale CG. A review of Reynolds stress models for turbulent shear flows. NASA CR-195054, 1995. Speziale CG, Gatski TB. Assessment of second-order closure models in turbulent shear flows. AIAA J 32: 21 13-21 15, 1994. Speziale CG, Thangam S. Analysis of an RNG-based turbulence model for separated flows. NASA CR-189600, 1992. Steffen CJ. A critical comparison of several low Reynolds number k-E turbulence models for flow over a backward-facing step. NASA TM-106173, 1993. Taulbee DB. An improved algebraic Reynolds stress model and corresponding nonlinear stress model. Phys Fluids A 4: 2555-2561, 1992. Thangam S, Speziale CG. Turbulent flow past a backward facing step: a critical evaluation of two-equation models. AIAA J 30: 1314-1320, 1992. Van Driest ER. On turbulent flow near a wall. J Aeronaut Sci 23: 1007-101 1, 1956. Wilcox DW. Dilatation-dissipation corrections for advanced turbulence models. AIAA J 30: 2639-2646, 1992. Wilcox DW. Turbulence Modeling for CFD. DCW Industries, Inc, 1993a. Wilcox DW. Application of low Reynolds number two-equation turbulence models to high Reynolds number flows. In: So RMC, Speziale CG, Launder BE, eds. Near-Wall Turbulent Flows. Elsevier, 1993b, pp 155-164. Yakhot V, Orszag SA. Renormalization group analysis of turbulence. I : Basic theory. J Sci Comput 1 : 3-51, 1986. Yoshizawa A. Statistical analysis of the deviation of the Reynolds stress from its eddy-viscosity representation. Phys Fluids 27 : 1377-1 387, 1984. Yoshizawa A. Statistical modeling of passive-scalar diffusion in turbulent shear flows. J Fluid Mech 195: 541-555, 1988. Youssef MS, Nagano Y, Tagawa M. A two-equation heat transfer model for predicting turbulent thermal fields under arbitrary wall thermal conditions. Int J Heat Mass Transfer 35: 3095-3104. 1992.

Turbulence Modeling

115

Zeman 0. Dilatational dissipation : the concept and application in modeling compressible mixing layers. Phys Fluids A 2: 178-188, 1990. Zhang HS, So RMC, Gatski TB, Speziale CG. A near-wall second-order closure for compressible turbulent flows. In: So RMC, Speziale CG, Launder BE, eds. Near-Wall Turbulent Flows. Elsevier, 1993, pp 209-21 8. Zukauskas A, Slanciauskas A, Karni J. Heat Transfer in Turbulent Fluid Flows. Hemisphere, 1987.

This page intentionally left blank

4 ~~~

Grid Generation .

Vijay K Garg A Y T CorporationlNASA Lewis Research Center. Clewland. Ohio

Philip C. E Jorgenson N A S A Lewis Research Cenrer. Cletdund. Ohio

4.1 Introduction Vijay K . Garg . . . . . . . . . . . . . . . 118 4.1.1 Metric Tensor and Physical Features of the Coordinate Transformation . . . . . . . . . . . . . . . . . . 120 4.1.2 Orthogonal and Conformal Coordinate Systems . . . . . . . 122 4.2 Grid Generation via Partial Differential Equation Solution V v a y K . Garg . . . . . . . . . . . . . . 4.2.1 Conformal Mapping: General Considerations . . . 4.2.2 Sequential Conformal Mapping . . . . . . . . 4.2.3 One-Step Conformal Mapping . . . . . . . . 4.2.4 Orthogonal Grid Generation . . . . . . . . . 4.2.5 Elliptic Grid Generation . . . . . . . . . . 4.2.6 Multiblock Grids . . . . . . . . . . . . . 4.3 Algebraic Grid Generation Vgay K . Garg . . . . . 4.3.1 One-Dimensional Stretching Functions . . . . . 4.3.2 Two-Boundary Technique . . . . . . . . . . 4.3.3 Multisurface Method . . . . . . . . . . . 4.3.4 Transfinite Interpolation . . . . . . . . . . 4.4 Adaptive Grid Generation Vijay K . Garg . . . . . 4.4.1 Method of Attraction and Repulsion . . . . . . 4.4.2 The Principle of Equidistribution . . . . . . . 4.4.3 Variational Approach . . . . . . . . . . . 4.4.4 Tension and Torsion Spring Approach . . . . .

. . . . . 123 . . . . . 123 . . . . . 125 . . . . . 128 . . . . . 130 . . . . . 133 . . . . . 136 . . . . . 140 . . . . . 140 . . . . . 141 . . . . . 143 . . . . . 145 . . . . . 147 . . . . . 147 . . . . . 148 . . . . . 150 . . . . . 151 117

118

Garg and Jorgenson

4.5 Unstructured Grid Generation: An Introduction Philip C . E. Jorgenson . . . . . . . . . . . 4.5.1 Introduction . . . . . . . . . . . . . . 4.5.2 Unstructured Grids . . . . . . . . . . . . 4.5.3 Triangular Grid Generation . . . . . . . . . 4.5.4 Triangulation by Point Insertion . . . . . . . 4.5.5 Data Structures . . . . . . . . . . . . . 4.5.6 Special Features . . . . . . . . . . . . . 4.5.7 Future Directions for Unstructured Grid Generation 4.6 Closure Vijay K . Gary . . . . . . . . . . . . Nomenclature . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . .

4.1

. . . . . . . . . . .

. . . . . . . . .

157 157 158 158 160 169 170 171 171 172 173

INTRODUCTION

The numerical solution of partial differential equations requires some discretization of the field into a collection of points or elemental volumes (cells). The differential equations are approximated as a set of finite difference or finite element or finite volume equations on this collection, and the resulting set of algebraic equations is then solved for the discrete values of the variables on this grid. The algebraic equations can be obtained on an organized or an unorganized distribution of points or cells. The former leads to a structured grid, while the latter leads to an unstructured grid. Though unstructured grids are increasingly in use these days, they cannot resolve the gradients near the walls, and are thus unable to provide heat transfer results. We will devote most of the chapter to the generation of structured grids, leaving the discussion on unstructured grid generation to the last section. The reader is referred to Thompson et al. (1982, 1985) for a comprehensive review of structured grid generation techniques, and to Choo (1995) and Carey (1997) for recent updates including unstructured grid generation. The organized discretization of the field has been handicapped by two problems. The first stems from the fact that most fields of interest are arbitrarily shaped regions, and thus accurate application of the boundary conditions requires that the discretization conform to the boundaries. This has led to the use of boundary-fitted coordinate systems, as we will see. The second problem in the numerical solution of partial differential equations is the resolution of grid points in regions of the field where very high changes

Grid Generation

119

in the solution occur. These high gradient regions are not known a priori, and therefore presetting of a fine mesh in these regions is out of question. Lack of prior information about these high gradients can render the grid useless since it is not able to resolve these high gradient regions satisfactorily. This makes the numerical simulation itself a waste since important physical phenomena do occur in the high gradient regions. Due to nonlinear phenomena associated with various physical processes such as boundary layer, turbulence, shock formation, and others, there is a tendency for the most important physical processes to occur in high gradient regions. These regions may or may not be associated with solid boundaries, and can also move in time. Thus, the problem of accurate resolution of high gradient regions is important not only from truncation error considerations but also from the physical point of view. The need to have an accurate physical simulation of these high gradient regions, and the lack of a priori information about these regions have led to the development of adaptive grid techniques. This technique automatically controls the grid size depending upon the driving function-usually the gradient of the solutionmaking the grid very fine in high gradient regions and relatively coarse in low gradient regions. The generation of grids for one-dimensional (1-D) problems is straightforward. There are many functions or other methods that can be used to generate a suitable 1-D grid. Moreover, the problem of complex boundaries does not arise in 1-D problems. That is why most work in structured grid generation has been done in two dimensions. Grid generation in three dimensions is very complicated and is generally achieved by stacking several two-dimensional (2-D) grids in the third direction. The general problem of grid generation is that of determining the mapping that transforms the grid points from the complex physical domain into the regular (generally rectangular in 2-D) computational domain. Such a mapping needs to satisfy several criteria, some of which are: 1. The mapping must be one-to-one. 2. The grid lines (surfaces in three dimensions) should be smooth to provide continuous transformation derivatives. 3. Grid points should be closely spaced in the physical domain where large numerical errors are expected. 4. Excessive grid skewness should be avoided. It has been shown (Raithby, 1976)that grid skewness exaggerates truncation errors.

For an arbitrary three-dimensional domain, it is not possible to generate an orthogonal grid. However, a grid that is almost orthogonal near the boundaries facilitates the application of boundary conditions. Although strict orthogonality is not necessary, the accuracy deteriorates if the departure

Garg and Jorgenson

120

from orthogonality is too large (Thompson et al., 1985). The implementation of turbulence models is more reliable with near-orthogonality at the boundary, since information on local boundary normals is usually required in such models. Algorithms based on the parabolized Navier-Stokes equations require that coordinate lines approximate the flow streamlines and the lines normal thereto, especially near solid boundaries. It is thus better in general if grid lines can be nearly orthogonal to the boundaries. Structured grid generation techniques can be classified into two categories. 1. Partial differential equation methods 2. Algebraic methods

In the former, the partial differential system may be elliptic, parabolic, or hyperbolic. Included in the elliptic systems are both the conformal and the quasi-conformal mappings, the former being orthogonal. Orthogonal systems, not necessarily conformal, may be generated from hyperbolic or elliptic systems. Some procedures are designed to produce nearly orthogonal coordinates. The algebraic procedures include simple normalization of boundaries, transfinite interpolation from boundary surfaces, the use of intermediate interpolating surfaces, and other related techniques. 4.1.1

Metric Tensor and Physical Features of the Coordinate Transformation

The generalized coordinates were introduced in section 1.11 when the governing equations were expressed in that coordinate system. In order to understand the following discussion better, let us link the generalized coordinates to the orthogonal and conformal coordinates. To do so, we introduce the metric tensor yij, which is related to the Jacobian matrix J in Eq. (1.95). We represent the physical domain by Cartesian coordinates .xi (=.x, y, z ) , i = 1, 3, and the computational domain by generalized coordinates t'(= 5, q, C), i = 1,3. The small distance As between two points in physical space can be written in terms of the coordinate displacements as AS'

1 Axk A.yk k = l

(4.11

The physical coordinate displacements Axk can be related to displacements in the generalized coordinates Ati by

Axk = 7 Ati ?Xk

(summation over i implied)

(4.2)

121

Grid Generation

Thus the small distance As related to generalized coordinates becomes 3

As2 = k = l = gij

Ati)(%

Ati A t j

Atj)

(summation over i a n d j implied)

(4.3)

where

The metric tensor g i j , discussed at length by Aris (1962, p. 142), relates the distance As to small changes in the generalized coordinates Ati. The metric tensor expressed in matrix form is related to the inverse Jacobian by

Taking determinants yields J g J 1 ’=2

(4.5)

The metric tensor g i j can be interpreted in terms of physical features of the computational grid. We will demonstrate this in two dimensions here. The three-dimensional forms are given by Kerlick and Klopfer (1982). According to Fig. 4.1, the grid cell area is given by Area

= g

)1/2

A t Aq

(4.6)

which, from ( 4 3 , gives a physical interpretation of the inverse Jacobian determinant. The physical orientation of the computational grid (tangent to a t-coordinate line) relative to the x-axis is given by the direction cosine (4.7) where xg = ax/a

-0.6

19 x 20 x 75. The computed potential flow pressure coefficients, along the outer boundary at 8 = 0", 90", and 180", and along the centerline, are shown in Fig. 5.6. The outer wall values at 8 = 90" and the centerline values are essentially identical and reflect the increase in cross-sectional area. The outer wall values at 8 = 0" and 180" have the typical shape for an S-duct, with higher pressure on the outside of the bend and lower pressure on the inside. The wiggles at the downstream end are probably due to the discontinuous slope in the cross section axes a and b, described previously. As noted earlier, they had no significant effect on the viscous solution. 5.3.5

Distortion Criteria

Before describing the viscous calculation, it is useful to discuss the method used to quantify the amount of distortion. The distortion descriptors are based on compressor face total pressure values that normally would be measured in an experiment by a standard 40-probe rake, (Fig. 5.7). To get analogous results from the CFD calculation, the computed total pressures

Inlet, Duct, and Nozzle Flows

Figure 5.7

199

Standard 40-probe compressor face rake.

were interpolated from the much denser computational grid to the probe locations of a 40-probe rake. For this study, a simplified stability assessment procedure supplied by the engine manufacturer was used. Four distortion descriptors are computed from the compressor face total pressure values, quantifying various aspects of the radial and circumferential distortion. These descriptors are then combined with empirical parameters, which are functions of corrected engine airflow rate, to define DLP(core) and DLP(fan), the distortion limit parameters for the core compressor and the fan tip. Both of these distortion limit parameters must be below 1.0 for stable engine operation.

5.3.6 Viscous Flow Solution/Analysis of Results The next step was to compute the viscous flow in the inlet without vortex generators. This was done for all three operating conditions listed in Table 5.1. In addition, since no experimental data was available to determine the boundary layer thickness at the throat (the initial station in the marching analysis), each operating condition was run with three different initial

200

Towne

Figure 5.8 Total pressure contours, high-speed cruise condition, without vortex generators.

boundary layer thicknesses-d/R = 0.025,0.05, and 0.10. The development of the flow through the inlet is illustrated in Fig. 5.8, in the form of computed total pressure contours at selected streamwise stations, for the high-speed cruise condition with 6/R = 0.05. The horseshoe-shaped pattern at the exit is typical of S-duct flows. The curved centerline causes transverse pressure gradients to be set up in the cross section, as shown by the potential flow results in Fig. 5.6. Pressure-driven secondary flow vortices appear, which drive the low-energy boundary layer flow to the bottom of the duct. The computed recoveries and distortion limit parameters are shown in Table 5.2 for all three operating conditions and initial boundary layer thicknesses. The total pressure recovery is satisfactory for all operating conditions and inlet boundary layer thicknesses. In addition, both DLP(fan) and DLP(core), the fan tip and core compressor distortion limit parameters, are well below the critical value of 1.0 at the best cruise and takeoff conditions, as is DLP(core) at the high-speed cruise condition. But, DLP(fan) at the high-speed cruise condition is clearly too high. This operating point was thus used to design the vortex generator system for the inlet. Based on earlier experience in the use of vortex generators in subsonic inlets (Anderson and Gibb, 1992; Anderson et al., 1992), a system was

201

Inlet,Duct, and Nozzle Flows Table 5.2

Distortion Limit Parameters without Vortex Generators

Operating condition

initial 6/R

High-speed cruise

Best cruise

Takeoff

Recovery ~~

~~~

0.025 0.05 0.10 0.025 0.05 0.10 0.025 0.05 0.10

0.982 0.979 0.972 0.985 0.982 0.978 0.992 0.990 0.988

DLP(fan) ~~~

DLP(core) ~~

2.156 2.510 2.300 0.238 0.259 0.264 0.087 0.107 0.137

0.008 0.076 0.307 0.024 0.105 0.279 0.002 0.003 0.037

designed for the Spirit inlet with 11 pairs of counterrotating generators distributed around the 360" cross section a short distance downstream of the throat, as shown schematically in Fig. 5.9. The design variable examined was the generator height h. Cases were run using R N S 3 D with h/R = 0.04 to 0.07 in increments of 0.005 for all three initial boundary layer thicknesses. The resulting values for the fan tip distortion limit parameter are shown in Fig. 5.10. Based on these results a generator height of h/R = 0.05

Figure 5.9

Vortex generator installation in the Spirit inlet.

Towne

202 3 .O

3a'

0.040

0.045

0.060

0.065

0.070

2.5

2.0

1.5

1.0

.C.

.2 0.5

0.0

0.035

0.050 0.055

0.075

Vortex generator height, h/R Figure 5.10

Effect of generator height on fan tip distortion limit parameter

was selected as the optimum choice, given the uncertainty in the actual boundary layer thickness. The computed total pressure contours for the h/R = 0.05 and 6,/R= 0.05 combination are shown in Fig. 5.11. By comparison with the results shown in Fig. 5.8, it can be seen that the effect of the vortex generators in this case is to split the large region of low total pressure at the compressor face into two smaller regions and to shift them slightly in the circumferential direction. Additional computations were also performed to confirm that the distortion levels at the other two operating conditions were still acceptable with vortex generators installed. The results for the three operating conditions and initial boundary layer thicknesses are summarized in Table 5.3. By comparison with Table 5.2, it can be seen that all the distortion levels were lowered by the use of vortex generators. The values for DLP(fan) are all below 1.0, the limit set for the candidate engine, but they are still uncomfortably high at the high-speed cruise operating point. The total pressure contours at the compressor face for this condition (see the exit station in Fig. 5.11) are actually very similar to the results for the best cruise condition (not shown). The high DLP(fan) values are a result of the large corrected weight flow value of 75.3 lb,,,/sec (34.1 kg/sec) at the high-speed cruise condition (Table 5.1). In the stability assessment procedure used in this study, the DLP(fan) values increase rapidly when the corrected engine airflow increases above 68 1bJsec (30.8 kg/sec).

Inlet, Duct, and Nozzle Flows

Figure 5.11 generators.

203

Total pressure contours, high-speed cruise condition, with vortex

After discussing these results with the inlet designers and the engine manufacturer, it was determined that a lower throat Mach number of 0.61 should have been used for these calculations. At this Mach number, the corrected weight flow is 71.8 lb,/sec (32.6 kg/sec). The high-speed cruise cases were therefore rerun with the lower throat Mach number, both with Table 5.3

Distortion Limit Parameters with Vortex Generators

Operating condition High-speed cruise

Best cruise

Takeoff

Initial 6/R

Recovery

DLP(fan)

DLP(core)

0.025 0.05 0.10 0.025 0.05 0.10 0.025 0.05 0.10

0.984 0.982 0.975 0.989 0.986 0.981 0.991 0.991 0.988

0.818 0.762 0.862 0.124 0.144 0.189 0.095 0.102 0.1 17

0.008 0.007 0.026 0.006 0.007 0.038 0.003 0.004 0.006

204

Towne

Table 5.4 Distortion Limit Parameters, High-speed Cruise, Mthr= 0.61

Without vgs

With vgs

Initial 6/R

Recovery

DLP(fan)

DLP(core)

0.025 0.05 0.10 0.025 0.05 0.10

0.985 0.981 0.976 0.987 0.985 0.979

0.448 0.541 0.564 0.259 0.271 0.301

0.007 0.052 0.218 0.007 0,006 0.045

and without vortex generators, and the resulting performance parameters are listed in Table 5.4. At the lower throat Mach number, the distortion levels are below 1.O even without vortex generators, although a conservative designer may feel they are still too high. With vortex generators, though, the levels are well below 1.0. Prior experience with R N S 3 D has shown that having a sufficiently dense mesh, especially in the streamwise direction, is required to get quantitatively accurate predictions of the development of secondary flow vortices. Some additional runs were therefore made for the high-speed cruise condition with vortex generators to investigate the effects of streamwise mesh density. One run was also made doubling the mesh in both cross-flow directions. The results are listed in Table 5.5. The initial boundary layer thickness 6 / R for these cases was 0.05. Using four times as many cross-section points increased the computed value of DLP(fan) only slightly. Going from 151 to 601 streamwise points, though, increased the value by over 50%. As the number of points was increased further, DLP(fan) dropped slightly and appeared to asympTable 5.5 Effect of Mesh Density on Distortion Limit Parameters Mesh

49 49 49 49 49

x x x x x

49 49 49 49 49

x 151

x 301 x 601 x 1201 x 2401

97 x 97 x 151

Recovery

DLP ( f an)

DLP(cor e)

0.985 0.982 0.979 0.980 0.981

0.271 0.344 0.411 0.379 0.363

0.006 0.004 0.004 0.004 0.004

0.983

0.310

0.023

Inlet, Duct, and Nozzle Flows

Figure 5.12

205

Effect of streamwise mesh density on DLP(fan).

totically approach a value of about 0.35, as shown more clearly in Fig. 5.12. This value is still well below the limit of 1.0 for the candidate engine. Based on these CFD results, Paragon, in consultation with the engine manufacturer, concluded that the Spirit inlet would meet the performance criteria for the candidate engine and that an experimental test program that had been planned could be eliminated with minimal risk. They plan to proceed directly to a flight test with an instrumented inlet installed on the new aircraft. Thus, the careful use of the CFD in this project has resulted in significant savings in both cost and time. 5.4

EXAMPLE-STRUT-JET

ENGINE

Propulsion systems for missiles, reconnaissance aircraft, and single-stage-toorbit vehicles must operate efficiently at flight conditions ranging from takeoff to hypersonic cruise. Because a specific propulsion cycle is more efficient at one flight condition than others, a new family of combined cycle

Towne

206

engines is being studied. These engines combine two or more different propulsion cycles into an integrated system for better overall performance throughout the flight envelope. One such system currently being studied at the NASA Lewis Research Center is the strut-jet. This engine is based on the rocket-based combined cycle (RBCC) concepts of Escher, Hyde, and Anderson (1995). It combines a high-specific-impulse low-thrust-to-weight air-breathing engine with a lowspecific-impulse high-thrust-to-weight rocket engine. From takeoff to high supersonic speeds (about Mach 3) the system operates as an air-augmented rocket. At approximately Mach 3 the rockets are shut down, and the system becomes a dual-mode ramjet. At very high Mach numbers (above about Mach 8) and high altitude, the air-breathing system may not provide adequate thrust, and the rockets would then be turned back on. Demonstration tests of the strut-jet engine were run at NASA Lewis in 1996. Thrust measurements were made at several fuel flow conditions. The measured experimental thrust, however, includes aerodynamic forces on various pieces of attached external hardware, such as the instrumentation and model support system, and therefore does not represent the true thrust of the propulsion system. The true thrust may thus be written as Tsys

Texp

Text

where TYsis the true internal thrust of the propulsion system, Kxp is the measured thrust in the experiment, and Tex, is the (negative) thrust due to external hardware. If the same experiment is run without fuel flow to the engine, we get where (KyJnf is the internal force on the propulsion system, and the measured force in the experiment. Subtracting, we can write

(Kxp)nfis

where AT = Texp - (TxJnfis the measured increment in thrust for a given fuel flow condition. The internal force (Tsys)nf cannot be measured experimentally because the force Text due to external hardware cannot be determined independently. ( TyJnfcan be computed using CFD, however, allowing the true thrust of the propulsion system to be determined. To compute this internal force, the NPARC code was used. NPARC is a multiblock Navier-Stokes code being developed and supported by the NPARC Alliance, a partnership between the NASA Lewis Research Center and the USAF Arnold Engineering Development Center (NPARC Alliance, 1994). It solves the Reynolds-averaged, unsteady compressible Navier-

Inlet, Duct, and Nozzle Flows

Figure 5.13 removed.

207

Rocket-based combined cycle engine: (a) sidewall removed; (b) cowl

Stokes equations in generalized nonorthogonal body-fitted coordinates. Several turbulence models are available in the code; for this application, the Chien low-Reynolds-number k-E model was used (Chien, 1982). Spatial derivatives in NPARC are represented using central difference formulas, and explicit boundary conditions are used. Jameson’s artificial dissipation model is used for stability, and to smooth pre- and postshock oscillations and to prevent odd-even point decoupling (Jameson, Schmidt, and Turkel, 1981). The equations are solved by marching in time using an AD1 algorithm derived using the Beam-Warming approximate factorization scheme (Beam and Warming, 1978). After the NPARC CFD calculations were completed, the solution was postprocessed to obtain the internal force ( TyJnf.Two calculation methods were used. The first was a simple momentum balance, subtracting the integrated momentum at the duct entrance from the integrated momentum at the exit. The momentum was computed by numerical integration over the computational grid. The second method integrated the pressure and skin friction forces on the internal surfaces of the configuration to obtain the internal force. Ideally, these two methods give identical results. However, several sources of error can contribute to a discrepency. Incomplete mass continuity and difficulty in calculating accurate skin friction are the two most common problems.

Towne

208

This methodology was first tested on a subscale model of the strut-jet engine that was tested in the NASA Lewis 1 x 1 ft (0.3 x 0.3 m) supersonic wind tunnel (Fernandez et al., 1996). A simplified schematic of the model is shown in Fig. 5.13. The strut-jet configuration that was analyzed consisted of a rectangular cross-sectioned inlet with swept leading-edge sidewalls. Two struts, also with swept leading edges, were installed in the inlet. In the actual engine test that will be run in the Hypersonic Test Facility, the rockets will be installed in the base of these struts. A precompression plate upstream of the inlet was used to simulate the effect of the vehicle forebody. The computational grid was created using GRIDGEN, a grid generation package widely used for C F D applications (Steinbrenner, Chawner, and Fouts, 1990). Six grid blocks were used, as listed in Table 5.6, with a total of 1,400,319 points. Note that the configuration is symmetric, and thus only half the duct was computed, from the sidewall to the center symmetry plane between the two struts. For these calculations, the freestream Mach number, static pressure, and static temperature were 6.0, 14.98 lb,/ftz (717.2 N/m2), and 93.13 R (51.74 K), respectively. The resulting Reynolds number was 3.80 x 106/ft (12.5 x 106/m). Convergence was achieved after the L , norm of the residual was reduced at least three orders of magnitude in each grid block, and when no discernable change in the static pressure and mass flow distributions were observed over at least 1000 iterations. Representative results are shown in Fig. 5.14a and b, where the computed and experimental static pressures distributions are plotted along the body and cowl centerlines. The internal force values computed using the two calculation methods were within 1.5% of each other, as shown in Table 5.7. Positive values indicate thrust, and negative values indicate drag. This example is an illustration of how C F D can be used to solve a problem that would be very dificult and expensive to solve any other way. Experimentally, there is no practical way to separate the true propulsion system thrust from the measured thrust at these hypersonic conditions. Table 5.6 Block no. 1 2 3 4 5 6

Grid Blocks for Strut-Jet Engine Computation Grid size

Description

x 76 x x 57 x x 57 x x 76 x x 95 x x 57 x

Inlet entrance center duct Center duct Side duct Inlet entrance side duct Forebody Combustor section

22 111 111 22 59 51

30 51 52 80 104

Inlet, Duct, and Nozzle Flows

Teble 5.7

209

Internal Force Balance for Strut-Jet Engine Momentum balance

Boundary

Inflow Spillage Outflow

Total

Momentum Clbt (NI1 -32.3999 ( - 144.122) 0.8108 (3.607) 28.3773 (126.228)

-3.2118 (-14.267)

Force integration Pressure “bf (NI1

Skin friction

Surface Body Cowl Strut Sidewall Base

- 1.3563 (-6.033) 0.0000 0.4575 (2.035) 0.0000 02566 (1.141)

-0.4740 -2.108) - 0.3226 - 1.435) -1.1703 - 5.206) - 0.6499 -2.891) 0.0000

Total

[lb,

- 3.2590 ( - 14.497)

(NI1

210

Towne 25

Experimental data

,O 15 4

O w

!i 2

E v)

10.0

12.5

15.0

20.0

17.5

22.5

25.0

27.5

30.0

Distance, x (inches) (a)

40 Experimental data 8

-$ 30 4

2a v)

10.0

12.5

15.0

17.5

20.0

22.5

25.0

27.5

30.0

Distance, x (inches) (b)

Figure 5.14 Pressure distribution in strut-jet engine: (a) along body centerline: (b) along cowl centerline.

Inlet, Duct, and Nozzle Flows

21 1

Using CFD, however, allows the true thrust of the propulsion system to be determined. 5.5 CURRENT STATUS AND FUTURE DIRECTIONS

Over the last 10 years or so, several papers and journal articles have appeared describing the status of CFD for applications. By their very nature, of course, publications like these become outdated fairly quickly. This section presents this author’s perception of the current status of C F D for inlet, duct, and nozzle applications and indicates possible future directions that would make C F D more useful in the real world (i.e., industry). Much of the material here has been influenced by the authors of two recent papers on the use of C F D in the aerospace industry (Cosner, 1994; Paynter, 1994). There are several issues, sometimes overlapping, that are inhibiting the widespread routine use of CFD, especially in the design environment. Some of these are modeling issues, resulting from our lack of understanding of some of the basic but complex flow physics in many real-world applications. Some are numerical issues, dealing with how the equations are solved. Others are more procedural in nature, related to how the various steps involved in a C F D analysis are currently being accomplished, and to how C F D codes are written, tested, and evaluated. 5.5.1 Modeling Issues For nonreacting flow through inlets, ducts, and nozzles, the principal flow modeling problems remaining today are the following. Turbulence

As noted earlier, a universal turbulence model that works well for all types

of flow does not yet exist. In general, guidelines based on experience must be used when choosing the model to use for a particular problem. This is an active area of research, and new turbulence models, or variations on existing ones, seem to be proposed weekly. This rapid growth makes it difficult to evaluate new models, however. The situation would improve with the development and acceptance of standards for software interfaces, validation, and documentation. (See below.) L aminar- Turbulent Transition

Our capability to predict laminar-turbulent transition is even less mature than that for fully turbulent flows. In many C F D codes the flow must be either fully laminar or fully turbulent. Those that are capable of predicting

212

Towne

transition generally use fairly crude models based on correlations with experimental data for simple flows. Better models are needed for use in computing realistic 3-D flows in engineering applications. See the recent papers by Simon (1993) and Simon and Ashpis (1996) for an overview of the current research in this area. Boundary Conditions C F D is typically used to study the flow through a component of a larger physical system, such as the inlet in a jet engine. At some types of boundaries, such as a simple no-slip solid wall or a supersonic inflow boundary, choosing appropriate boundary conditions is fairly straightforward. For many other types of boundaries, the situation is more complicated. The conditions specified at the outflow boundary of an inlet, for example, must properly model the influence of the compressor on the flow in the inlet. Porous wall boundary conditions are normally used to represent the flow through a bleed region in a supersonic inlet. These specialized boundary conditions must also be able to model unsteady interactions between components, such as the reflection at the compressor face of a disturbance in the inlet flow. Additional research is required to develop satisfactory boundary conditions for specialized applications like these. 5.5.2 Numerical Issues

The numerical algorithms being used in modern C F D codes to solve the governing equations are generally pretty fast. While faster algorithms are always desirable, other numerical issues are also of critical importance. Computational Platform

The computational power available to the C F D user has increased dramatically over the last 10-15 years. In the not-too-distant past, CFD codes were almost always run on large mainframe computers, but today NavierStokes analyses for relatively simple 2-D problems can be run on desktop PCs and workstations. Even some 3-D problems are being run on midrange to high-end workstations. Parallel processing software has been developed that allows C F D codes to use multiple processors, either on a single computer with multiple CPUs or on a cluster of computers, with each processor computing a part of the problem. This rapid and continual growth in the capability of the hardware has in many respects been the determining factor in the growth of CFD. The computer speed and memory that is available to the CFD user influences, for example, the size of the grid and the sophistication of the turbulence model. As the hardware continues to improve, C F D simulations will also continue to improve.

Inlet, Duct, and Nozzle Flows

213

Other advancements are possible in CFD, besides those related to the raw speed and memory of the computer. New solution algorithms that are designed specifically to take advantage of parallel processing capabilities should be investigated. Faster algorithms may also be developed by taking advantage of other features present in a specific type of computational architecture. The disadvantage to this, and it is a big one, may be lack of portability between platforms. For long-term use in a design environment, the trade-off is probably not worthwhile. Unsteady Flows

While many C F D codes are at least theoretically able to compute unsteady flows, not much emphasis has been placed on their numerical accuracy. Interest in unsteady flows is growing, however, and this area will see increasing activity in the future.

5.5.3 Procedural Issues Another reason, perhaps the main reason, that C F D is not more widely used in routine design work is that the process is still too difficult and time-consuming, especially for non-CFD experts. There are several, sometimes interrelated, factors involved. Ease of Use

The computer codes used in CFD, from the preprocessors used to define the geometry and generate the grid, through the postprocessors used to analyze the results, need to be made simpler to use. Until fairly recently C F D was basically a research area, with much effort being put into the development of faster and more accurate solution algorithms. Ease of use for the nonexpert user was not a high priority for the code developer. However, the situation is changing. Solution algorithms are now pretty good, as noted above, and ease of use is becoming more important. Part of the solution will require closer coupling between the various steps in the C F D process, and the development of various standards will help, as discussed below. There are other things that should also be done, however, to make the individual steps in the solution process easier. For complex configurations, grid generation is currently one of the more time-consuming steps. In particular, setting up the various blocks for a multiblock analysis and linking the grid blocks can be very labor intensive (Cosner, 1994). Automating this step as much as possible would be very beneficial. Cosner suggests an expert system type of approach, in which the key geometric features would be identified, and, along with the expected flow conditions, used as input to a system that would recommend the layout of the grid blocks, plus the grid size and grid point distribution

214

Towne

within the blocks. A corollary to this idea is the use of adaptive grid techniques, in which the grid points are automatically redistributed to resolve high-gradient regions as the flow is being computed. Adaptive gridding is not a new idea, and has already been demonstrated for a variety of problems. However, it has not yet become a standard feature in most C F D analysis systems, perhaps because it requires close coupling between the grid generator and the flow solver. Adaptive gridding has the potential, though, to make the initial grid generation step much easier, to eliminate much of the manual iteration that is now sometimes necessary between the grid generator and flow solver, and to allow the best possible solution for a given number of grid points. The flow solvers themselves can also be made easier to use. Some C F D codes are overly sensitive to things like nonorthogonal and nonsmooth meshes, the time step size, and the choice of artificial viscosity parameters. Getting good results (or in the most extreme cases, any results at all) from a C F D code may require “tweaking” the input until the “correct” value or combination of values is found. More robust solution algorithms, that are less sensitive to mesh and input irregularities, would help. Another improvement would be the development of an intelligent user interface, that could be used to help set up the input for a particular problem and to check it for inconsistencies. Postprocessing systems, while generally very good, can still be improved. As noted earlier, with the interactive 3-D graphics packages available today, results may be displayed and examined in almost any form imaginable. While this can be tremendously useful, for the most part the user must visually examine the computed results. More automated methods should be developed to identify key flow features and problem areas. The capability to perform solution quality checks, similar to the grid quality checks already available in some grid generation codes, is also needed. These automated postprocessing capabilities, once available, should be used to examine the flow field as it is being computed, and recommend changes to the grid and/or input parameters where appropriate. Standardization

The various steps in the solution process have historically been separate elements. As a result, too much of the time required to solve a problem is spent in between the elements, converting the output from one step to the input for the next step. In addition, there is often too much iteration required between steps (e.g., “change the grid and recompute”). Closer coupling between the various steps is needed, so that the entire solution process from the geometry specification through the analysis of the results becomes as seamless as possible. Using CFD for design requires, almost by

Inlet, Duct, and Nozzle Flows

215

definition, the ability to easily change the geometry and determine the effect of that change on the flow. To accomplish this, standards need to be developed and accepted by the CFD community for the interfaces between the various steps. For example, the wide variety of CAD packages in use for geometry specification have resulted in a variety of formats for the CAD output, many of them not directly readable by popular grid generation programs. CFD flow solvers read grid files in a variety of formats, and there is no universal standard for the interface requirements and boundary conditions to be used between blocks in a multiblock grid. The PLOT3D format (Walatka et al., 1990) has become a de facto standard for the output from CFD codes and can be read by a variety of postprocessors. Unfortunately, this format does not include all the information necessary to fully represent some computed results, such as turbulence data. A variety of more complete formats have been or are being developed, such as the NASA-IGES standard for CAD output (Blake et al., 1991) and the interface standards from the NASA-funded Complex Geometry NavierStokes (CGNS) project. Since there is no “governing body” in CFD, however, the development of a single accepted standard for the interface between each step in the solution process is unlikely, at least in the near future. Instead, several “standards” will probably coexist. Code developers should therefore strive to support directly as wide a variety of the proposed formats as possible, both for input and output. In addition, generalized interface routines should be developed to convert data between a variety of standard formats. Besides the need for standard data formats between steps in the solution process, standard interfaces are needed between modules within the individual codes. This is especially true for the CFD flow solver itself, where standard subprogram interfaces would make the development and testing of new technology, such as improved turbulence models, much easier. The CGNS project is addressing this issue also. Finally, research into multidisciplinary methods, such as a C F D analysis coupled with an elastic structure analysis, is increasing. For these methods to ultimately be useful in the real world, standards are required for sharing data between the multiple analyses involved. Validation

C F D code validation has been the subject of much interest in recent years (e.g., Marvin, 1993; Mehta, 1995; Aeschliman, Oberkampf, and Blottner, 1995). While various terms, such as verification, certification, and validation, have been used to describe different aspects of the process, it basically refers to determining how well a C F D code is able to simulate reality.

216

Towne

In order to determine the strengths and weaknesses of a C F D code, cases should be run for a variety of geometric configurations and over a range of flow parameters. Computed results should be compared with benchmarkquality experimental data, well-accepted computational results, and/or analytic solutions. If C F D is to become an accepted tool for design, code validation must be emphasized. The code developer must demonstrate that his or her code is able to simulate reality accurately enough and quickly enough to be relied upon in a design environment. Starting as far back as 1968, various organizations have developed databases containing standard sets of experimental data to be used for C F D validation for various types of flow (e.g., Coles and Hirst, 1968; AGARD, 1988; Settles and Dodson, 1991). As our capability to predict more complex flows increases, the need for high-quality experimental validation data for those flows also increases. Data are now needed for high(i.e., flight level) Reynolds-number turbulent flows, low-Reynolds-number transitional flows, and unsteady flows. These data would be especially useful for evaluating newly proposed turbulence models. Documen ta f io n

C F D codes are notoriously poorly documented. For many C F D codes the documentation, if it exists at all, consists only of a user’s guide describing the input and output, with a few examples. It generally does not include a detailed description of the code itself, showing exactly how the various physical and numerical models involved have been implemented. These details are often not even described in comments within the code itself. Without this information, even a knowledgeable C F D researcher will have difficulty modifying the code to test hypotheses about the cause of any disagreement with experimental data in a validation study. Papers presenting applications of CFD are also often poorly documented. In addition to describing the problem and the CFD method that was used, they should include at least brief descriptions of the turbulence model, any artificia1 viscosity that was used, the grid size and distribution, the boundary conditions, and, for iterative methods, the iteration method and convergence history. Without these details, it is difficult or impossible to assess the significance of the computed results. ACKNOWLEDGMENTS

The author would like to acknowledge and thank his co-workers at NASA Lewis Research Center and in the NPARC Alliance for their valuable contributions to this chapter. Special thanks go to Bernie Anderson and Julie Dudek, who did most of the work on the Paragon Spirit inlet calculations;

Inlet, Duct, and Nozzle Flows

217

to Jim DeBonis and Shaye Yungster, who supplied the material on the strut-jet engine; and to Ray Cosner and Jerry Paynter, who provided insight into the difficulties involved in using CFD for real-world design in an industrial environment. NOMENCLATURE

a, b DLP(core) DLP(fan) h Mm Mthr P Pr R Re Re, t 9

Texp 9 T e x t

Tsys

semiaxes in Spirit inlet cross section core compressor distortion limit parameter fan tip distortion limit parameter vortex generator height freestream and throat Mach numbers for Spirit inlet static pressure Prandtl number equivalent throat radius in Spirit inlet Reynolds number Reynolds number based on throat conditions for Spirit inlet streamwise marching parameter in R N S 3 D code measured thrust, thrust due to external hardware, and true system thrust corrected engine airflow centerline coordinates for Spirit inlet

Greek

boundary layer thickness Subscripts

nf r W

value at no-fuel-flow condition reference quantity wall value

REFERENCES Aeschliman DP, Oberkampf WL, Blottner FG. A proposed methodology for computational fluid dynamics code verification, calibration, and validation. 16th International Congress on Instrumentation in Aerospace Simulation Facilities, Wright-Patterson AFB, OH, July 18-21 1995. AGARD. Validation of computational fluid dynamics. Vol. 1. Symposium papers and round table discussion. AGARD-CP-437, 1988. Ames Research Staff. Equations, tables, and charts for compressible flow. NACA report 1 135, 1953.

218

Towne

Anderson BH. The aerodynamic characteristics of vortex ingestion for the F/A- 18 inlet duct. AIAA paper 91-0130, 1991. Anderson BH. Three-dimensional viscous design methodology of supersonic inlet systems for advanced technology aircraft. In: Murthy SNB, Paynter GC, eds. Numerical Methods for Engine-Airframe Integration. New York : AIAA. 1986. Anderson BH, Gibb J. Application of computational fluid dynamics to the study of flow control for the management of inlet distortion. AIAA paper 92-3177, 1992. Anderson BH, Huang PS, Paschal WA, Cavatorta E. A study on vortex flow control of inlet distortion in the re-engined 727-100 centre inlet duct using computational fluid dynamics. AIAA paper 92-0152, 1992. Anderson BH, Towne CE. Application of computational fluid dynamics to inlets. In: Goldsmith EL, Seddon J, eds. Practical Intake Aerodynamic Design. Oxford: Blackwell and New York: AIAA, 1993. Anderson JD. Modern Compressible Flow with Historical Perspective. New York: McGraw-Hill, 1982. Beam RM, Warming RF. An implicit factored scheme for the compressible NavierStokes equations. AIAA J 16: 393-402, 1978. Blake MW, Kerr PA, Thorp SA, Chou JJ. NASA geometry data exchange specification for computational fluid dynamics. NASA RP 1338, 1991. Briley WR, McDonald H. Analysis and computation of viscous subsonic primary and secondary flows. AIAA paper 79-1453, 1979. Briley WR, McDonald H. Three-dimensional viscous flows with large secondary velocity. J Fluid Mech 144: 47-77, 1984. Chien K-Y. Predictions of channel and boundary-layer flows with a low-Reynoldsnumber turbulence model. AIAA J 20: 33-38, 1982. Coles DE, Hirst EA. Computation of turbulent boundary layers-1968 AFOSRIFP-Stanford Conference. Vols I, 11. Stanford University, 1968. Cosner RR. Issues in aerospace application of CFD analysis. AIAA paper 94-0464, 1994. Escher WJD, Hyde EH, Anderson DM. A user’s primer for comparative assessments of all-rocket and rocket-based combined-cycle propulsion systems for advanced earth-to-orbit space transport applications. AIAA paper 95-2474, 1995. Fernandez R, Trefny CJ, Thomas SR, Bulman M. Parametric data from a wind tunnel test on a rocket based combined cycle engine inlet. NASA TM 107181. 1996. Jameson A, Schmidt W, Turkel E. Numerical solutions of the Euler equations by finite volume methods using Runge-Kutta time-stepping schemes. AIAA paper 81-1259, 1981. Kunik WG. Application of a computational model for vortex generators in subsonic internal flows. AIAA paper 86-1458 (also NASA TM 87327), 1986. Levy R, Briley WR, McDonald H. Viscous primary/secondary flow analysis for use with nonorthogonal coordinate systems. AIAA paper 83-0556, 1983.

Inlet, Duct, and Nozzle Flows

219

Levy R, McDonald H, Briley WR, Kreskovsky JP. A three-dimensional turbulent compressible subsonic duct flow analysis for use with constructed coordinate systems. AIAA paper 80-1398, 1980. Marvin JG. Dryden lectureship in research, a perspective on CFD validation. AIAA paper 93-0002, 1993. Mehta UB. Guide to credible computational fluid dynamics simulations. AIAA paper 95-2225, 1995. NPARC Alliance. A User’s Guide to NPARC Version 2.0, 1994. Numbers KE. Survey of CFD applications for high speed inlets. WL-TR-94-3 131, 1994. Paynter GC. CFD status for supersonic inlet design support. AIAA paper 94-0465, 1994. Povinelli LA, Towne CE. Viscous analyses for flow through subsonic and supersonic intakes. NASA TM 88831 (Prepared for the AGARD Propulsion and Energetics Panel Meeting on Engine Response to Distorted Inflow Conditions, Munich, Germany, Sept. 8-9, 1986), 1986. Settles GS, Dodson LJ. Hypersonic shock/boundary-layer interaction database. NASA CR 177577, 1991. Simon FF. A research program for improving heat transfer prediction for the laminar to turbulent transition region of turbine vanes/blades. NASA TM 106278, 1993. Simon FF, Ashpis DE. Progress in modeling of laminar to turbulent transition on turbine vanes and blades. NASA TM 107180, 1996. Steinbrenner JP, Chawner JR, Fouts CR.The GRIDGEN 3D multiple block grid generation system. WRDC-TR-90-3022, 1990. Towne CE. Computation of viscous flow in curved ducts and comparison with experimental data. AIAA paper 84-0531 (also NASA TM 83548), 1984. Towne CE, Anderson BH. Numerical simulation of flows in curved diffusers with cross-sectional transitioning using a three-dimensional viscous analysis. AIAA paper 8 1-0003 (also NASA TM 8 1672), 1981. Towne CE, Povinelli LA, Kunik WG, Muramoto KK, Hughes CE. Analytical modeling of circuit aerodynamics in the new NASA Lewis altitude wind tunnel. AIAA paper 85-0380 (also NASA TM 86912), 1985. Towne CE, Schum EF. Application of computational fluid dynamics to complex inlet ducts. AIAA paper 85-1213 (also NASA TM 87060), 1985. Tsai T, Levy R. Duct flows with swirl. AIAA paper 87-0247, 1987. Vakili A, Wu JM, Hingst WR, Towne CE. Comparison of experimental and computational compressible flow in an S-duct. AIAA paper 84-0033, 1984. Walatka PP, Buning PG, Pierce L, Elson PA. PLOT3D User’s Manual. NASA TM 10 1067, 1990.

This page intentionally left blank

Turbine Flows: The Impact of Unsteadiness .

Om P Sharma. Daniel J . Dorney.* Seyt Tanrikut. and Ron-Ho Ni Pratt & Whitney Aircraf. East Hartford. Connecticut

6.1 Introduction . . . . . . . . . . . . . . . . . 6.2 Unsteady Flow Effects in Turbines: Experimental Data . . 6.2.1 Upstream Row-Induced Effects . . . . . . . . 6.2.2 Upstream Stage-Induced Effects . . . . . . . . 6.3 Unsteady Flow Effects in Turbines: Numerical Simulations 6.3.1 Effect of Upstream Wakes on Rotor Secondary Flows 6.3.2 Hot Streak Migration through a li-Stage Turbine .

. . . . . 221

. . . . 226 . . . . 226 . . . . 234 . . . . 237 . . . . 242 . . . . 245 6.4 Implications of Results . . . . . . . . . . . . . . . . . . 252 6.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . 253 References . . . . . . . . . . . . . . . . . . . . . . 254

6.1

. . . . . .

INTRODUCTION

Although the importance of periodic unsteadiness. induced by the relative movements of adjacent airfoil rows in turbomachinery. has long been

*Current affiliation: G M I Engineering & Management Institute. Flint. Michigan . 22 1

Sharma et al.

222

acknowledged by design engineers, the unavailability of established prediction methods has precluded an explicit impact on the hardware design. The effects of periodic unsteadiness, until recently, were accounted for through empiricism in correlations and criteria used in design procedures. Application of these procedures often yield nonoptimal designs requiring expensive and time-consuming development programs. Work has been done over the past 15 years to develop more rigorous procedures to account for these unsteady flow effects as discussed below (Sharma et al., 1994). Experimental programs have been conducted to investigate the impact of periodic unsteadiness on the loss and heat load generation mechanisms in turbines. Highlights from these experiments indicate that losses and heat load in an unsteady flow environment are larger than those measured for the same airfoils in a steady flow environment, as shown in Fig. 6.1 (Hodson, 1983; Blair et al., 1988; Doorley et al., 1986; Sharma et al., 1988). Simple models (Doorley et al., 1986; Sharma et al., 1990; Speidel, 1957) are available to account for the effects of upstream wake-induced unsteadiness on the performance of the downstream airfoil. The model proposed in Sharma et al. (1988) showed that the change in the profile losses of an airfoil in an unsteady environment can be related to the “reduced frequency,” as shown in Fig. 6.2. Reduced frequency in this figure is defined as a ratio of the “flow change period” (relative speed/pitch of the upstream airfoil row) to the “flow interaction period” (axial velocity/axial chord of the downstream airfoil). In addition to providing a good estimate of profile losses for an embedded row of a multistage turbine, the correlation in Fig. 6.2 can also be used to calculate time-averaged boundary layer properties CALCUUlIONS

0 005

TUABUCLNr TRANSIlIONAl

0 ROTOR

0CASCADE

0 CASCAOE

o 0

(a)

: 10

NORMALIZED STREAMWISE DISTANCE

(b)

NORMALIZED STREAMWISE DISTANCE

Figure 6.1 Measured streamwise distribution of time-averaged (a) Stanton number of Hair et al. (1988) and Sharma et al. (1988) and (b) boundary layer thickness of Hodson (1983) show larger values in an unsteady environment than in a steady cascade configuration.

Turbine Flows

NOAMAClZED

Loss

223

0.6

Y - Y,

0.4

+ HOOSON f 1984)

oo % B

AEOUCEO FAEOUENCY

(fl =

Figure 6.2 Additional time-averaged loss generated due to the unsteadiness induced by upstream wakes can be related to the frequency (Sharma et al., 1988).

including heat transfer as shown in Fig. 6.3. These calculations were conducted by assuming an increased level of intermittency factor in the airfoil suction surface boundary layer; the net increase in the intermittency factor is deduced from the correlation in Fig. 6.2. Different models have been proposed (Scholz, 1977; Mayle and Dullenkopf, 1989; Hodson et al., 1992) to account for the effects of periodic unsteadiness on the airfoil boundary layer characteristics through the modification of the transition behavior, but results from these models are not found to be much different than those in Fig. 6.3. All of these models predict increases in both the loss and the heat load for downstream airfoils in the unsteady environment, provided the airfoils have attached boundary layers for steady incoming flows. In situations such as low-Reynolds-

HOOSON'S ROTOR TEST (1984)

BLAIR ET AL 11988.a.bt ROTOR TEST

o oos

OOWr

NORMALIZE0 Ooo2 MOMENTUM

LOSS

THICKNESS oQ)l

-CALCULATIONS 0

DATA CALCULArlONS

DATA

SHARMA ET AC (19881

0001

NORMALIZED STREAMWISE OISTANCE

Figure 6.3 Sharma model (Sharma et al., 1988) yields good estimates of timeaveraged (a) momentum loss thickness and (b) Stanton numbers from data acquired in an unsteady flow environment.

Sharma et al.

224

number operating conditions, where airfoils have extended regions of separated laminar flows (Fig. 6.4), interaction from upstream wakes can result in lower losses in an unsteady environment relative to the steady flow operating conditions. No reliable model is available to predict the behavior of airfoils operating at low Reynolds numbers. It should be pointed out here that measured airfoil surface static pressure data in the above experiments could be well predicted by using steady flow codes, indicating that periodic unsteadiness has an insignificant effect on the airfoil loadings. Results in Fig. 6.3 were obtained by using timeaveraged experimental data for the airfoil loadings. This implies that if airfoil loadings do not get affected by unsteadiness, fair estimates of timeaveraged profile losses and heat loads can be obtained. Experience indicates that periodic unsteadiness has a relatively small effect on time-averaged loadings for turbines operating with moderate axial spacing between adjacent airfoils and operating at subsonic flow speeds. Recent calculations (Rangwalla et al., 1991), conducted by using an unsteady Reynolds-averaged Navier-Stokes (RANS) code for the mean section of the first stage of a transonic turbine, indicated that the timeaveraged loadings on the upstream stator are strongly affected by the axial gap between the stator and the rotor. These results, plotted in Fig. 6.5, clearly show that the time-averaged diffusion on the stator is significantly

0.15

Stalled Airfoil

Airfoil With Attached Boundary Layers

Distance Loss e

Coefficient

Pressure ,Predicted Losses By UsingBoundary [

Distance -0 0

-6

REYNOLOS NUMBER (millions) Figure 6.4 High loss levels measured for airfoils at low Reynolds numbers. Boundary layer separation is evident on airfoil surfaces.

Turbine Flows

225

Figure 6.5 Diffusion on the upstream airfoil affected by the axial gap between rows. This effect is not accounted for in classical "design systems."

reduced as the axial gap between the airfoil rows is increased. Experimental verification of this numerical result is needed to enhance confidence in the predictive capabilities of the C F D codes. The predicted time-averaged loadings for the above-mentioned stator at the largest axial gaps were found to be quite close to those predicted by using steady C F D codes currently used in the Pratt & Whitney (P&W) design system. The results discussed show limitations of the current prediction methods in accounting for the effects of periodic unsteadiness on airfoil loadings, which in turn affect performance and heat transfer coefficient distributions. The main focus of this chapter is to identify flow situations which can be adequately predicted by C F D codes used in the current design process and those flow situations which cannot be well predicted. This chapter also points out research work required to enhance the predictive capabilities of the current CFD codes to model these flow situations. In the next section, the results obtained by analyzing data from a number of experimental turbine programs are described in order to provide an improved understanding of the flow physics. The impact of upstream wakes, hot streaks, and secondary flow vortices on the performance, heat load, and flow distribution through downstream airfoil rows are also discussed. Advancements made in the flow simulation capabilities for multistage turbines through the use of unsteady C F D codes are discussed in section 6.3. It is shown that these unsteady codes provide realistic predictions of

Sharma et al.

226

flow features through multistage turbines by using relatively simple models for the effects of turbulent viscosity. Implications of the information are discussed in section 6.4. A clarification on situations where steady and unsteady CFD codes are needed in the design process is given. Conclusions are outlined in section 6.5.

6.2

UNSTEADY FLOW EFFECTS IN TURBINES: EXPERIMENTAL DATA

All airfoil rows in turbines encounter spatial and temporal flow distortions generated by upstream airfoil rows and combustors, resulting in unsteady flows. Flow in the first-stage rotor passage is influenced by temporal distortions from adjacent stators. These temporal distortions consist of wakes, vortices, and entropic (temperature) disturbances from upstream stators, and potential (acoustic) waves from the upstream and downstream stators. Flow in the second-stage stator is affected not only by the temporal distortions generated by the adjacent rotor airfoils but also by the spatial distortions generated by the first-stage stator. The effects of temporal distortions, termed “upstream row-induced effects,” are discussed in section 6.2.1. The effects of spatial distortions, termed “upstream stage-induced effects,” are discussed in section 6.2.2. 6.2.1

Upstream Row-Induced Effects

The circumferential variations in the velocity field downstream of the firststage stators in turbines are normally generated by the drag on the airfoil and endwall surfaces which cause reduced velocities and increased turbulence. For airfoil rows downstream of a combustor, high-velocity jets are found to exist due to large circumferential gradients in temperature. The effects of these upstream velocity variations can be simply illustrated through the use of velocity triangles (Fig. 6.6, Butler et al., 1989; Kerrebrock and Mikolajczak, 1970). This figure shows that the low-velocity fluid has a slip velocity toward the suction side of the downstream airfoil (for the compressor, the slip velocity is toward the pressure side) indicating that the high-turbulence, low-velocity fluid from the upstream airfoil wake will migrate toward the suction side of the airfoil. In a similar manner, highvelocity (high-temperature) fluid will migrate toward the pressure side of the downstream airfoil. This preferential migration of fluid particles has three effects:

Turbine Flows LOW FLOW COEFFICIENT

227 HIGH FLOW COEFFICIENT

STATOA

COEFFICIENT AXIAL V E L O C I N

WHEEL SPEED

W R E U T I W E VELOCITV V ABSOLUTE VELOCITY U WHEELSPEED

Figure 6.6 Rotor inlet gas temperature distortion causes large variation in blade incidence angle. Simple calculations conducted for hot-to-cold temperature ratio of 1.7 indicate angle variations of 12" and 40" for typical high- and low-flow-coefficient configurations, respectively.

1. Alterations in the boundary layer characteristics of the airfoil

through its effect on the transition process. This effect is reasonably well accounted for in the turbine design process as outlined above. 2. Variations in the secondary flow generation in downstream passages. 3. Redistribution of the stagnation enthalpy (temperature). Detailed discussions of the second and the third effects are given below. Effect of Upstream Wakes on Secondary Flows

Total pressure loss data (Sharma et al., 1985) obtained by using highresponse probes in the United Technologies Research Center (UTRC) largescale rotating rig (LSRR) for the rotor as it passes through the upstream stator flow field, are shown in Fig. 6.7. This figure illustrates contours of the relative total pressure coefficient upstream and downstream of the rotor passage. In this figure, the residence time of the fluid particles in the rotor passage is accounted for in such a manner that the exit flow field corresponds to the given inlet flow field. Large variations in the rotor exit flow structures are seen in the figure for three different inlet conditions. These inlet conditions correspond to different positions of the upstream stators

Sharma et al.

228 CPTA = RELATIVE TOTAL PRESSURE - REFERENCE PRESSURE OYNAMlC HEAO BASE0 O N WHEEL SPEED AT MID-SPAN

TIP SECONOARY FLOW VORTEX

UNSTEAOY INTERACTION

TIP LEAKAGE VORTEX

UNSTEAOY I N TEA ACT10 N - 2

F I R S T STATOR ROOT SECONDARY FLOW V O R T E X

CoNToU" I

CPTA 2.6

51)

Figure 6.7 Total pressure loss contours and gap-averaged profiles at inlet and exit of the rotor in relative frame indicating the influence of unsteadiness.

relative to the rotor passage. When the inlet flow is circumferentially uniform due to the rotor passage being positioned between two adjacent stator wakes, the exit flow field shows three distinct vortices (Fig. 6.7a). The vortices are due to the hub and tip secondary flows, and the tip leakage effects. Without the tip leakage vortex, the flow field in Fig. 6.7a is similar to the one expected for this airfoil in a steady cascade environment. As the

Turbine Flows

229 A WALL TEMPERATURE

J-:

O F

400

- - ---

0.0

SUCTION SIDE I

PRESSURE SIDE

Figure 6.8 Hotter pressure sides indicated in turbine rotors. Temperature difference between pressure and suction surfaces of 250°F observed.

rotor passes through the upstream stator flow field, the tip leakage vortex shows little variation (Fig. 6.7a-c), indicating that the leakage phenomenon is not influenced by upstream circumferential distortions. The hub secondary flow vortex shows the largest variation, transforming from a distinct structure in Fig. 6.7a to a diffused structure in Fig. 6.7b, and becoming almost nonexistent in Fig. 6 . 7 ~ The . overall variation in the size of the tip secondary flow vortex is smaller than that of the hub vortex but larger than the leakage vortex (Fig. 6.7~).This indicates that the secondary flow generation mechanisms, especially at the hub, are strongly influenced by the upstream circumferential distortions such as wakes. The periodic oscillation in the size and strength of the secondary flow vortices observed in this experimental investigation equate to almost 40% variation in the secondary flow losses for the rotor passage. These data indicate that there is a potential to reduce secondary flow losses by manipulating the secondary flow vortices and increasing the unsteadiness in the rotor by increasing the number of upstream stator airfoils. Effect of Upstream Temperature Streaks

Heat loads on turbine rotors are also affected by the migration of hot and cold air from the upstream stator and the combustor. Experimental data

230

Sharma et al.

Turbine Flows

23 1

acquired in an engine indicate that pressure and suction sides of an airfoil can operate in different temperature regimes; the results plotted in Fig. 6.8 show that the differences can be on the order of 250‘F. Extensive work has been done during the past 10 years to highlight the physical mechanisms responsible for this segregation of hot and cold air in turbine rotors and it is discussed here. Results from an experimental investigation, conducted to quantify the influence of combustor induced hot streaks on the segregation of hot and cold air in turbine rotors, are discussed. In this investigation, the experimental data were acquired in the UTRC LSRR by introducing temperature streaks at inlet to the first-stage stator (Fig. 6.9). Two types of temperature profiles were generated upstream of the first-stage stator: 1. A hot streak generated in a circular pipe to yield temperature profiles both in the radial and the circumferential directions; some of the results from this investigation were reported in Sharma et al. (1990) and Butler et al. (1989). Data were acquired for two different flow coefficients by restaggering the stator and by increasing the speed of the rotor to maintain design incidence angle on the rotor airfoil. 2. A hot streak generated with a rectangular nozzle to yield a radially uniform profile that had temperature gradients in the circumferential direction. This experiment was conducted for the lower flow coeficient only. The hot air in these experiments was seeded with carbon dioxide (CO,) and the migration path of the hot streak through the turbine was deduced using static pressure taps and “sniffing” techniques as discussed in Butler et al. (1989). The temperature patterns at the exit of the first-stage stator for this test are given in Fig. 6.9b. The temperature patterns follow the Munk and Prim principle (Munk and Prim, 1947) and indicate little mixing in the stator passage. Measured concentrations of CO, on the rotor airfoil surfaces from the nozzle (rectangular hot streak) experiment are shown in Fig. 6.10. This figure shows higher levels of COz on the rotor airfoil pressure

Figure 6.9 (a) Schematic of the experimental apparatus used to simulate the redistribution of a hot streak in a turbine rotor (Butler et al., 1989). (b) Contour plots of normalized CO, concentration downstream of the first-stage stator in the UTRC LSRR obtained with circular and rectangular hot streaks. High values indicate high temperature. (c) Spanwise distribution of normalized CO, concentration profiles (indicators of temperature) measured in the rotor frame from the circular and rectangular hot streaks.

232

Sharma et al.

Figure 6.10 Larger time-averaged CO, concentration (temperature) measured on the pressure side of the rotor airfoil relative to the suction side indicates the segregation of hot and cold air.

Figure 6.11 Effect of inlet temperature distortion can be reduced by increasing the flow coefficient.

Turbine Flows

233

Figure 6.12 Original alternate turbopump design (ATD) turbine test article (TTA) measured efficiency contours.

side than on the suction side. These results, obtained with a radially uniform inlet profile (Fig. 6.9c), clearly demonstrate that the segregation of the hot and cold air in turbine rotors is mainly driven by two-dimensional mechanisms. Experimental data acquired on the rotor with a circular incoming hot streak are shown in Fig. 6.11 for two flow coefficients. These

234

:: - A

* 3

Sharma et al.

Vane Passiton

Circ L o c a t m

6' 40

5'20'

4'0'

2'40'

1'2J

6'40'

Figure 6.13

ATD TTA first vane clocking positions

data indicate that the rotor pressure side temperature is higher for the lower flow coefficient configuration. Extrapolation of these data to an actual engine environment indicate that the pressure side of first-stage rotors can operate at temperatures between 100-700°F higher than the suction side. These temperature differences can cause significant durability problems for airfoils and endwalls. Large amounts of cooling air are required to accommodate these temperature levels, resulting in reduced efhiency of the cycle and increased specific fuel consumption of the engine.

6.2.2

Upstream Stage-Induced Effects

Experimental data show substantial variations in flow quantities measured downstream of a stage at stations where the second-stage airfoils are normally located. Indexing of second-stage airfoils relative to these incoming distortions can have a significant impact on the overall performance of the machine as discussed below. The performance of the alternate turbopump development (ATD) turbine test article was evaluated at NASA Marshal1 Space Flight Center during 1991 (Gaddis et al., 1992). Measured turbine efficiencies downstream of the second stage (Fig. 6.12) showed a 2-cycle pattern on top of a 54-cycle pattern, the latter corresponding to the second-stage stator airfoil count.

Turbine Flows

Figure 6.14

235

Measured efficiency contours for “clockable” turbine configuration.

The two-cycle pattern equated to a &OS% in the overall turbine efficiency, and was found to be due to the interaction between the first- and secondstage stators (i.e., dependent on where the wake fluid from the first-stage stator impinges on the second-stage stator). With airfoil counts for the firstand the second-stage stators of 52 and 54, respectively, a 2-cycle pattern will exist over the full annulus, as indicated in Fig. 6.12. Experiments were conducted to assess the impact of indexing on the overall turbine performance and to establish whether this concept can be exploited for engine applications. Hardware was built to allow indexing of the two stators (Fig. 6.13). Both temperature and torque were measured to define the efficiency. Flow-field data were acquired over the full annulus. Results from these experiments (Huber et al., 1995) clearly indicate that the

236

Sharma et al.

Figure 6.15 (a) Efficiency as a function of clocking position at midspan clearly indicates a minimum and a maximum. (b) Overall efficiency as a function of clocking position also indicates a minimum and a maximum.

performance of the turbine can be optimized by appropriately indexing first- and second-stage stator airfoils (Figs 6.14 and 6.15). Experimental data acquired over the range of incidence angles also showed large changes in performance due to indexing effects.

Turbine Flows

237

Figure 6.15 Continued. (c) Both thermodynamic and mechanical measurement of turbine efficiency in the ATD TTA confirm clocking effects and magnitude of the variations.

The above results indicate that a knowledge of the shapes and locations of wakes from upstream airfoils is required to optimize the design of downstream stage airfoils. 6.3

UNSTEADY FLOW EFFECTS IN TURBINES: NUMERICAL SIMULATIONS

Significant progress has been made over the past 20 years in developing flow prediction systems based on CFD codes. The application of these CFD codes in the turbine design process is shown in Fig. 6.16, where the interrelationships of the C F D codes with conventional design procedures are illustrated. The main contribution from CFD has come through the use of multistage flow analyses. Three different approaches are available to compute time-averaged flows through multistage machines. 1. In the first approach, flow through each airfoil row in the machine is calculated for specified circumferentially uniform inlet and average exit boundary conditions. These boundary conditions are initially obtained from predicted or " data-matched'' streamline curvature methods. Subsequently, these boundary conditions are deduced from circumferentially

Sharma et al.

238

Figure 6.16

Current turbine design practices are based on steady CFD codes.

averaged mean flow quantities obtained from computations for adjacent airfoil rows. This approach (Ni and Bogoian, 1989; Denton, 1990; Dawes, 1990, 1990b), known as the “mixing plane” approach, was extensively used during the 1980s to establish spanwise distributions of airfoil loadings and flow profiles. An example of this is shown in Fig. 6.17. This approach relies on the solution of equations governing the conservation of mass, momentum, and energy, while the impact of periodic unsteadiness represented by apparent-stress ”-like terms is neglected. The effect of periodic unsteadiness has, therefore, not been accounted for in this approach. 2. The second approach, termed the “average passage” approach, was developed (Adamczyk, 1985) to accurately simulate time-averaged flows through multistage machines. The effects of adjacent airfoil rows in this approach are accounted for through the use of body forces and “apparent stresses.’’ Reliable models are not yet available to account for circumferential variations of “apparent stresses. These are currently assumed to be constant in the circumferential direction. The average passage approach has the potential to yield more accurate estimates of flow through multistage machines at off-design conditions than the mixing plane approach. Significant work, however, is needed to develop physics-based models to “

”

Turbine Flows

239

ELEVATION VIEW O f A TWO-STAGE HIGH PRESSURE TURBINE RIG.

COMPUTATIONAL BOUNDARIES FOR THE MULTI-STAGE FLOW SIMULATION.

COMPUTATIONAL MESH USED FOR THE W O STAGE TURBINE FLOW SIMULATION.

Figure 6.17 Three-dimensional steady multistage Euler code of Ni and Bogoian (1989) provides accurate estimates of airfoil loadings and total pressure as well as total temperature profiles in multistage turbines.

account for the radial and circumferential variations of apparent stresses and to enhance the predictive capabilities of codes based on the averagepassage approach. In a number of situations the flow field may need to be computed over more than one airfoil passage in the machine; this implies a

240

Sharma et al.

Turbine Flows

241

flow simulation over the most pertinent circumferential dimension that includes multiple airfoil passages. This requires minor modifications to mixing-plane and average-passage approaches. This point is further discussed in section 6.3.2. 3. Unsteady flow computations in the third approach utilize either the Euler (Ni et al., 1989, 1990) or RANS (Rai, 1989; Rai and Madavan, 1988; Rao and Delaney, 1990; Gundy-Burlet, 1991; Rhie et al., 1995; Hall, 1997; Weberand and Steinert, 1997) equations. These computations, conducted for actual airfoil counts, can yield very accurate results for the timeaveraged flows. Simulations of actual airfoil counts in a multistage environment can require tremendous computational resources. The Euler codes, with approximate airfoil counts and with relatively coarse grids, have been extensively used in conducting design optimization studies. More recently the use of RANS codes has been incorporated into the design process. Numerical experiments conducted with unsteady RANS codes are providing information which indicates that their application to the design process needs to be accelerated as discussed below. One of the cornerstones of the design process is the ability to predict airfoil surface static pressure distributions which provide information about work, losses, and heat loads. Once the pressure distribution is known, design criteria and boundary layer calculation methods are used to select high-performing airfoil sections. The main reason why Euler flow solvers were so easy to incorporate into the design process was that these methods produced reliable predictions of airfoil surface static pressure distributions in cascades and in multistage rotating rig environments, thus gaining the confidence of design engineers. Initial positive results (Huber and Ni, 1989; Huber et al., 1985) from the application of these codes encouraged turbine design engineers to look for further improvements in durability and performance through the use of more advanced codes. Unsteady flow simulations of the UTRC LSRR model turbine have provided insight into the effects of upstream wakes and hot streaks on flow mechanics. These simulations (Ni et al., 1989; Takahashi and Ni, 1991) were conducted for single- and 1)-stage configurations using a three-dimensional unsteady multistage Euler code. The airfoil count in these simulations was

Figure 6.18 (a) Computational mesh of Ni and Bogoian (1989) for the UTRC large scale rotating rig (LSRR). (b) Predicted time-averaged pressure distributions from the 3D Euler code of Ni and Bogoian (1989) show good agreement with measured data for the LSRR first stage. Both steady and unsteady multistage codes yield similar results for loadings.

242

Sharma et al.

three first-stage stators, four first-stage rotors, and four second-stage stators (instead of the experimental ratio of 22 first-stage stators, 28 first-stage rotors, and 28 second-stage stators) in order to contain computer requirements. The first-stage stators were scaled to maintain mass flow and pitchto-chord ratio. Two sets of simulations were conducted, one with a uniform upstream total temperature profile (single stage) to simulate the flow conditions from Sharma et al. (1985) and the other with an upstream temperature streak (1 *-stage configuration) to simulate flow conditions from Butler et al. (1989). A wall shear stress model (Denton, 1986) was used in the code to simulate the viscous flow effects. The tip leakage flow was not modeled in these calculations. The results from these simulations are discussed below. 6.3.1 Effect of Upstream Wakes on Rotor Secondary Flows

The computational mesh used in the single-stage simulation for the UTRC LSRR (Ni et al., 1989) is shown in Fig. 6.18a. A total of 70,000 grid points were used in the axial, radial, and tangential directions to discretize the flow field. The time-accurate flow solver and an interpolation method at the blade-row interfaces were used to obtain the unsteady periodic solution in time. Convergence was obtained at about 20,000 time steps or six rotorpassing cycles (a cycle is defined as the rotor through a distance equal to one stator airfoil pitch) and required about 10 CPU hours on a CRAYX M P computer for each simulation. The predicted time-averaged loadings on the stator and the rotor airfoils from both simulations are compared to the experimental data in Fig. 6.18b for three spanwise locations. The predicted results are shown to be in good agreement with the experimental data. It should also be pointed out that these solutions are almost identical to those obtained by the steady multistage code (Ni and Bogoian, 1989), indicating that unsteadiness has a weak impact on the airfoil loadings in the UTRC LSRR. The effects of upstream wakes on the secondary flow generation in the downstream rotor can be deduced through the review of unsteady total pressure contours downstream of the rotor from the simulation conducted with a uniform upstream total temperature profile. Contour plots of the computed instantaneous relative total pressure coefficients downstream of rotor, together with the 3-stator/4-rotor configuration, are shown in Fig. 6.19. These results indicate the unsteadiness has a strong effect on the rotor flow field. Organized flow structures pointed out in this figure indicate the existence of different secondary flow vortices in each rotor airfoil passage. The secondary flow vortices in the tip region are similar in all four rotor passages, which indicates that circumferential distortions generated by the upstream stator have relatively little effect on the flow in the tip region.

Turbine Flows

243

TION A ~ J ~ A T ~ ~ Relative Total Pressure Loss at Rotor Exit

Figure 6.19 Relative total pressure loss at the exit of the rotor: (a) location of data plane, and (b) periodic disappearance of the root secondary flow vortex is in agreement with the experimental data.

244

Sharma et al.

Figure 6.20 Four snapshots in time of an isotherm of one hot jet in a I$-stage flow simulation.

Turbine Flows

245

Significant passage-to-passage variations in the flow structures are observed in the root region. In particular, there are substantial reductions in the strength of the root secondary flow vortex as the rotor moves past the stator airfoils. This periodic disappearance of the root secondary flow vortex is in excellent agreement with experimental data (Sharma et al., 1985), as shown in Fig. 6.19. The comparison indicates that the unsteady flow simulation using a Euler code successfully predicted the unsteady and distorted flow features observed in the experimental data with primitive modeling of viscous effects. Results from this numerical study clearly demonstrate that Euler codes can be used to obtain first-order effects of flow unsteadiness in turbines.

6.3.2 Hot Streak Migration through a If-Stage Turbine

A three-dimensional hot streak simulation was conducted by using one hot streak, three first-stage stators, four first-stage rotors, and four second-stage stators in the geometric model. Four views of the isotherms associated with the hot streak as it migrates through the turbine are shown in Fig. 6.20. One hot streak is located in every third first stator passage; only one is shown in the figure for clarity. A complex interaction of the hot streak with the rotor and the second-stage stator is suggested by this figure. The hot streak is injected at the inlet midway between two first-stage stators and it is found to convect through the first-stage stator passage with only minor changes due to area variations. The hot streak next enters the rotor passage, where it is chopped up by the passing rotor airfoils into discrete eddies. The high-temperature eddies are convected into the second-stage stator passages, where they are further broken up. The calculated time-averaged temperature distributions in four rotor passages are shown in Fig. 6.21. Note that each of the four rotor passages have identical time-averaged solutions because each rotor sees, over a periodic cycle, the identical inlet and exit boundary conditions. The hot gas tends to migrate toward the rotor pressure side (temperature segregation), and the rotor passage secondary flow transports the hot gas radially over the pressure surface and then over the endwalls (three-dimensional convection). Simultaneously, the hot gas on the rotor suction side appears to lift off the surface with increasing axial distance. An important observation from this figure is that maximum time-averaged temperatures downstream of the rotor leading edge are higher than the time-averaged maximum gas temperatures forward of the leading edge. These higher-thaninlet time-averaged temperatures are an indicator of the long residence time of the segregated hot gas in the rotor passage.

246

Sharma et al.

Figure 6.21 Time-averaged temperature in four rotor passages shows the hot jet segregation and migration.

A comparison of the temperature distribution in the rotor passage obtained by time-averaging the results of the unsteady simulations to those obtained from steady multistage (mixing-plane) simulations are shown in Fig. 6.22 in the form of contour plots at specified axial locations. This figure clearly shows that the maximum temperature from the steady multistage code is significantly lower than that obtained from the time-averaged results of the unsteady simulation. These results demonstrate that the timeaveraged temperature on the rotor airfoil surfaces is strongly affected by periodic unsteadiness induced by combustor generated hot streaks. Conventional design procedures rely on axisymmetric rotor inlet temperatures to define the heat load on rotor airfoils; this has historically resulted in the underestimation of airfoil pressure surface temperatures and inaccurate estimates of cooling air requirements. The present work indicates that an unsteady Euler code can be used to provide a more accurate estimate of the gas temperature near the airfoil surface, which should result in a better cooling air estimate in the design process. Two numerical experiments were conducted (McGrath et al., 1994) for the UTRC 1;-stage LSRR turbine to assess the degree of complexity

Turbine Flows

247

Figure 6.22 The maximum temperature from a steady solution is significantly lower than that from a time-averaged unsteady solution.

needed to model the effects of hot streak migration using steady multistage Euler codes. Since the flow in this problem is periodic over three first-stage stator passages, the simulation needed to be conducted using an averageperiod approach by utilizing three first-stage stators, four first-stage rotors, and four second-stage stators. In addition, the effects of periodic unsteadiness were accounted for through the use of apparent stresses. Distributions of apparent stresses were computed in the entire computational domain from the unsteady flow simulations discussed above. Circumferentially averaged values of these stresses, termed " axisymmetric apparent stresses ", were used in the first numerical experiment, while the three-dimensional distribution of stresses (nonaxisymmetric) were used in the second experiment. Results from these numerical experiments are shown in Fig. 6.23 in the form of contours of relative total temperature at various axial stations in the rotor passage. These results clearly show that axisymmetric stresses are insufficient to explain the segregation of hot and cold air in rotor passages. Results obtained by using nonaxisymmetric stresses are, however, in excellent agreement with those obtained from unsteady simulations (Fig. 6.21).

248

Sharma et al.

Figure 6.23 Segregation of hot and cold air is not predicted by an "averagepassage" code. Relative total temperature contours in the rotor passage are in the same view as given in Fig. 6.22.

Time-averaged temperatures in the four second-stage stator passages are shown in Fig. 6.24. A number of observations can be made from the numerical results shown in this figure: 1. Hot gas in the second-stage stator passage is found to be confined to a small region of the entire 4-stator flow solution domain. This is in contrast to the results obtained for the rotor passages, which had identical time-averaged temperature fields. 2. The maximum time-averaged temperature levels in the secondstage stator passage are significantly higher than in rotor passage. This is mainly because the hot gas in the second stator passage is

Turbine Flows

249

Figure 6.24 Time-averaged temperature in the four second-stage stator passages shows that hot gas is confined to a small region.

confined to a small region of the entire 4-stator flow solution regime. 3. The maximum downstream time-averaged temperatures are not higher than the maximum time-averaged temperatures at the second-stage stator inlet. This is opposite of what was observed in the time-averaged rotor flow field and indicates that hot gas is not lingering for a significant period of time on the second-stage stator surfaces. 4. There are no obvious signs of temperature segregation in the second stator passage; the hot gas does not have a tendency of preferential migration to either the pressure or the suction surface. Figure 6.24 also shows that the hot streak has been split as it reaches the leading edge of a second stator airfoil. The hot gas that splits to the pressure side in passage 2 stays attached to the pressure surface, migrating radially along the surface toward the endwalls with increasing axial distance. This radial migration of the hot gas is similar to that observed on the rotor pressure side (Fig. 6.21) and indicates transport with classical secondary flows generated in the stator passage. The hot gas that splits toward the

250

Sharma et al.

Figure 6.25 Time-averaged radial velocity contours show that there are four vortices in the second-stage stator passage and two vortices in the rotor passage.

suction side (passage 1 of Fig. 6.24) stays attached to the suction surface where it spreads radially toward both the inner and outer endwalls with increasing axial distance. This radial spreading on the suction side is opposite to the behavior observed on the rotor (Fig. 6.21), where the hot

Figure 6.26 Streamlines on the rotor suction side converge to midspan at the trailing edge, whereas on the second-stage stator suction side the streamlines diverge toward the endwalls.

Turbine Flows

Figure 6.27 stator.

251

Computed temperature distributions in the UTRC LSRR second-stage

gas on the rotor suction side migrated radially toward the midspan, transported by the secondary flow endwall vortices generated in the rotor passage. The suction surface hot gas in Fig. 6.24 is apparently being transported radially toward both endwalls by vortices near the suction side, for which analogous vortices did not appear in the rotor passage. The second stator passages apparently contain four primary vortices as shown by the radial velocity component contours in Fig. 6.25, which are taken at a planar cut in passage 1 near the airfoil trailing edge. The mechanism believed to be generating the two extra suction side vortices is the interaction of vortices generated in the rotor passage with the downstream second-stage stator. Figure 6.25 also shows the radial velocity component at the trailing-edge plane of the upstream rotor. There appears to be only

Sharma et al.

252

two primary vortices in the rotor passage, which are the classical counterrotating endwall vortices generated in the rotor passage. To further illustrate the differences between the rotor and second-stage stator secondary flows, the streamlines on the rotor and second-stage stator airfoil surfaces are shown in Fig. 6.26. The pressure side of both the rotor and second-stage stator show streamlines bending toward the endwalls with increasing axial distance from the leading edge, consistent with the existence of classical endwall secondary flow vortices on the pressure side. However, on the suction side the streamlines for the rotor bend toward the midspan region and the streamlines for the second-stage stator bend toward the endwalls. This observation is consistent with the existence of two primary vortices in the rotor and four in the second-stage stator. Figure 6.27 shows total temperature contours at five axial planes in the second-stage stator passage obtained from 1. Time-averaged results from the unsteady Euler code 2, Steady multistage (mixing-plane) code 3. Steady multistage (average-period) code where the effects of periodic unsteadiness are accounted for through nonaxisymmetric distributions of the apparent stress.

The figure illustrates that the results obtained from the average-period approach are almost identical with those obtained from the unsteady code. Results obtained from the mixing-plane approach, however, are much different than those obtained from the unsteady code. This comparison indicates that current steady multistage codes, which solve for the flow through an average passage using either the mixing-plane or the average-passage approach, are insufficient to describe flow features having circumferential length scales larger than one airfoil passage. An average-period approach is required to describe these flows; in addition, the effects of periodic unsteadiness need to be accounted for through three-dimensional distributions of apparent stresses. 6.4

IMPLICATIONS OF RESULTS

Results from both experimental data and unsteady numerical simulations indicate that time-averaged loadings at subsonic Mach numbers on airfoil surfaces (and endwalls), with moderate axial gaps between adjacent airfoil rows, are not influenced by unsteadiness even in the presence of hot (temperature) streaks. Airfoil loadings are, however, affected by potential waves in turbines with small axial gaps between adjacent airfoil rows. The nonreflecting boundary conditions suggested in Giles (1988) allow this effect

Turbine Flows

253

to be accurately accounted for in multistage steady CFD codes with the mixing-plane approach. The pressure side of turbine rotors operate at higher temperatures than the suction side in the presence of circumferentially nonuniform temperature profiles. Currently available steady multistage CFD codes do not predict this phenomenon. Unsteady multistage Euler and Navier-Stokes codes can predict this flow phenomenon, and the computational resources have recently become available to allow the application of these codes in the turbine design process. Unsteady Euler codes by themselves, however, are not sufficient to provide accurate estimates of heat loads since viscous regions on airfoil surfaces also contribute toward establishing magnitudes of heat loads on airfoil surfaces. A viable technique may be to utilize a multistage steady RANS code based on the average-period approach discussed in section 6.3.2. The effects of periodic unsteadiness in this code may be accounted for through the use of apparent stresses computed from numerical simulations conducted by using the unsteady Euler code. Further work is needed to demonstrate the accuracy of utilizing this technique in the turbine design process. Circumferential distortions generated by viscous flow mechanisms in first-stage airfoil passages have a significant effect on the performance of second-stage airfoils, as discussed in section 6.2.2. Significant improvements in the performance of multistage turbines may be achieved by indexing firstand second-stage airfoils. An improved physical understanding of the loss generation mechanisms, however, is needed to allow its routine application in the design process. Numerical simulations utilizing unsteady and,/'or average-period RANS codes should be able to provide this insight.

6.5

CONCLUSIONS

The following conclusions are drawn from the above discussions. 1. Three-dimensional steady multistage flow prediction codes provide accurate estimates of loadings for airfoil rows even in the presence of periodic unsteadiness and temperature distortions. 2. Turbine inlet circumferential temperature distortions result in hotter pressure sides and colder suction sides for rotor airfoils. This phenomenon is not predicted by the steady multistage codes which are currently used in turbine design procedures. Unsteady multistage Euler codes together with RANS codes based on the averageperiod approach are needed to provide accurate estimates of heat loads in a realistic turbine flow environment.

Sharma et al.

254

3. The eficiency of multistage turbines can be improved by indexing first- and second-stage airfoils. Improved understanding of the loss generation mechanisms, however, is needed to apply these concepts in a routine manner. 4. Continued development of unsteady RANS codes is needed to provide accurate estimates of the heat loads and losses in realistic turbine environments.

REFERENCES Adamczyk JJ. Model equation for simulating flows in multistage turbomachinery. ASME paper #85-GT-226, 1985. Blair MF, Dring RP, Joslyn HD. The effects of turbulence and stator-rotor interactions on turbine heat transfer. Part I. Design operating conditions. ASME paper # 88-GT-125,1988. Butler TL, Sharma OP, Joslyn HD, Dring RP. Redistribution of an inlet temperature distortion in an axial flow turbine stage. AIAA J Propulsion Power 5 : 64-71, 1989. Dawes WN. A comparison of zero and one equation turbulence modeling for turbomachinery calculations. ASME paper # 90-GT-303, 1990a. Dawes WN. Towards improved throughflow capability: the use of 3D viscous flow solvers in multistage environment. ASME paper # 90-GT-18, 1990. Denton J. Calculation of three-dimensional viscous flows through multistage turbine. ASME paper #90-GT-19, 1990. Denton JD. The use of a distributed body force to simulate viscous flow in 3D flow calculations. ASME paper # 86-GT- 144, 1986. Doorley DJ, Oldfield MLG, Scrivener CTJ. Wake passing in a turbine rotor cascade. AGARD CP-390, paper No. 7, Bergen, Norway. Gaddis SW, Hudson ST, Johnson PD. Cold flow testing of the space shuttle main engine alternate turbopump development high pressure fuel turbine model. ASM E paper # 92-GT-280, 1992. Giles MB. Stator-rotor interaction in a transonic turbine. AIAA paper # 88-3093, 1988. Gundy-Burlet, KL. Computations of unsteady multistage compressors flows in a workstation environment. ASME paper #91-GT-336, 1991. Hall EJ. Aerodynamic modeling of multistage compressor flowfields. Part I : Analysis of rotor/stator/rotor aerodynamic interaction. ASME paper # 97GT-344, 1997. Hall EJ. Aerodynamic modeling of multistage compressor flowfields. Part 2: modeling deterministic stresses. ASME paper # 97-GT-345, 1997. Hodson HP. The development of unsteady boundary layers in the rotor of an axialflow turbine. AGARD Proceedings No. 351, Viscous Effects in Turbomachines, 1983.

Turbine Flows

255

Hodson HP, Addison JS, Shepherdson CA. Models for unsteady wake-induced transition in axial turbomachines. J Phys I11 France 2: 545-574, 1992. Huber FW, Rowey RJ, Ni R-H. Application of 3D flow computation to gas turbine aerodynamic design. AIAA-85-1216, 1985. Huber FW, Ni R-H. Application for a multistage 3D Euler solver to the design of turbines for advanced propulsion system. AIAA paper # 89-2578, 1989. Huber FW, Johnson PD, Sharma OP, Staubach JB, Gaddis SW. Experimental investigation of vane wake clocking effects on turbine performance. ASME paper #95-GT-27, 1995; to appear in ASME J Turbomachinery. Kerrebrock JL, Mikolajczak AA. Intra stator transport of rotor wakes and its effect on compressor performance. ASME J Eng Power Oct : 359-370, 1970. Mayle RE, Dullenkopf K. A theory for wake-induced transition. ASME paper # 89GT-57, 1989. McGrath D, Sharma OP, Ni R-H, Takahashi RT, Stetson GM, Staubach JB. Accurate simulations of flow through multistage turbomachines by using 3D steady CFD codes. Private communication, 1994. Munk M, Prim RC. On the multiplicity of steady gas flows having the same streamline pattern. Proc Nat Acad Sci USA 33: 1947. Ni R-H, Bogoian JC. Prediction of 3D multistage turbine flow field using a multiple-grid Euler solver. AIAA paper # 89-0203, 1989. Ni R-H, Sharma OP. Using 3D Euler flow simulations to assess effects of periodic unsteady flow through turbines. AIAA paper # 90-2357, 1990. Ni R-H, Sharma OP, Takahashi R, Bogoian J, 3D unsteady flow simulation through a turbine stage. Paper presented at the 1989 Australian Aeronautical Conference-Research and Technology-The Next Decade, Melbourne, Australia, Oct., 1989. Rai MM. Three dimensional Navier-Stokes simulations of turbine rotor-stator interactions. AIAA J Propulsion Power 5: 307-319, 1989. Rai MM, Madavan, NK. Multi airfoil Navier Stokes simulations of turbine rotorstator interaction. AIAA paper 88-0361, 1988. Rao K, Delaney R. Investigation of unsteady flow through transonic turbine stage. Part I : Analysis. AIAA paper 90-2408, 1990. Rangwalla AA, Madavan NK, Johnson PD. Application of unsteady Navier Stokes solver to transonic turbine design. AIAA paper # 9 1-2468, 1991. Rhie CM, Gleixner AJ, Spear DA, Fischberg CJ, Zacharias RM. Development and application of a multistage Navier-Stokes solver. Part I: Multistage modeling using body forces and deterministic stresses. ASME paper # 95-GT-342, 1995. Scholz N. Aerodynamics of cascades. AGARD-AG-220, 1977. Sharma OP, Butler TL, Joslyn HD, Dring RP. Three-dimensional unsteady flow in an axial flow turbine. AIAA J Propulsion Power 1: no. 1, 1985. Sharma OP, Ni R-H, Tanrikut S. Unsteady flows in turbines-impact on unsteadiness. AGARD Lecture Series Turbomachinery Design Using CFD. Paper 195-5, May-June, 1994. Sharma OP, Pickett GF, Ni R-H. Assessment of unsteady flows in turbines. ASME paper # 90-GT- 150, 1990.

Sharma et al.

256

Sharma OP, Renaud E, Butler TL, Milsaps K, Dring RP, Joslyn HD. Rotor-stator interaction in multistage axial turbines. A I A A paper 88-3013, 1988. Speidel L. Beeinllussung der laminaren Grezschicht durch periodische Strorungen der Zustromung. Z Flugwiss 9: 5, 1957. Takahashi R, Ni R-H. Unsteady hot streak simulation through I f stage turbine. AIAA paper # 9 1-3382, 1991. Weber A, Steinert W. Design, optimization and analysis of a high-turning transonic tandem compressor cascade. ASME paper # 97-GT-412, 1997.

7 Numerical Modeling of Materials Processing and Manufacturing Systems Y ogesh J aluria Rutgers U nitvrsitj*. N e w Brumtrick . N e w Jersrj.

7.1 Introduction . . . . . . . . . . . . . . . . . . . 7.2 Governing Equations . . . . . . . . . . . . . . 7.2.1 General Equations . . . . . . . . . . . . . 7.2.2 Buoyancy Effects . . . . . . . . . . . . . . 7.2.3 Viscous Dissipation . . . . . . . . . . . . . 7.2.4 Processes with Phase Change . . . . . . . . . . 7.2.5 Numerical Solution . . . . . . . . . . . . . 7.3 Numerical Approximation of Boundary and Initial Conditions 7.3.1 General Conditions . . . . . . . . . . . . . 7.3.2 Free Surfaces and Openings . . . . . . . . . . 7.3.3 Other Conditions . . . . . . . . . . . . . . 7.3.4 Phase Change . . . . . . . . . . . . . . .

. . . 258 . . . . 262 . . . . 262 . . . . 263 . . . . 264 . . . . 264 . . . . 265 . . . . 268 . . . . 268 . . . . 270 . . . . 270 . . . . 271 7.4 Governing Parameters and Simplifications for Numerical Simulation . . 273 7.4.1 Common Simplifications . . . . . . . . . . . . . . . 273 7.4.2 Approximations and Transformations . . . . . . . . . . . 274 7.5 Numerical Coupling due to Material Properties . . . . . . . . . 277 7.5.1 Variable Properties . . . . . . . . . . . . . . . . . 277 7.5.2 Viscosity Variation . . . . . . . . . . . . . . . . . 278 257

258

Jaluria

7.5.3 Other Aspects . . . . . . . . . . . . . 7.6 Additional Considerations and Typical Numerical Results 7.6.1 Basic Aspects . . . . . . . . . . . . . 7.6.2 Material Property Variation . . . . . . . . 7.6.3 Moving Boundary . . . . . . . . . . . 7.6.4 Solidification . . . . . . . . . . . . . 7.6.5 Conjugate Transport . . . . . . . . . . . 7.6.6 System Simulation . . . . . . . . . . . 7.7 Concluding Remarks . . . . . . . . . . . . . Nomenclature . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . .

7.1

. . . . . . 280 . . . . . . 28 1 . . . . . . 282 . . . . . . 283 . . . . . . 285 . . . . . . 287 . . . . . . 290 . . . . . . 293

. . . . . . 295 296 . . . . . . . . . . 297

INTRODUCTION

In recent years, there has been a considerable interest and research activity in manufacturing and materials processing. This is mainly due to increased international competition and the critical need to optimize existing manufacturing systems and processes, improve product quality, reduce costs, develop new processes, and produce a variety of new materials. Materials processing generally refers to procedures and techniques that modify and combine given materials to obtain desired characteristics in the product, whereas manufacturing refers to the overall area of mass production of useful items. An improved understanding of the physical mechanisms underlying materials processing is needed for significant advancements in this area. In addition, it is necessary to determine the dependence of material characteristics on the process and operating conditions so that these may be modified for improving product quality and reducing costs. Mathematical modeling of the relevant thermal processes is the first step in this direction. However, the governing equations are usually very complicated, and numerical modeling and simulation, with validation by analytical or experimental results, are needed for most practical systems and processes. In a wide variety of manufacturing processes, heat transfer and fluid flow considerations play a very important role. These include processes such as casting, crystal growing, hot rolling, optical fiber drawing, soldering, welding, gas cutting, plastic injection molding and extrusion, metal forming, and heat treatment. Table 7.1 presents different types of materials processing operations along with a few important examples. Such manufac-

Modeling of Manufacturing Systems

259

Table 7.1 Different Types of Materials Processing Operations, Along with Examples of Commonly Used Processes Processes 1. Processes with phase change

2. Heat treatment 3. Forming operations

4. Cutting

5. Bonding processes 6. Plastic processing 7. Other processes

Examples Casting, continuous casting, crystal growing, drying Annealing, hardening, tempering, surface treatment, curing, baking Hot rolling, wire drawing, metal forming, extrusion, forging Laser and gas cutting, fluid jet cutting, grinding, machining Soldering, welding, explosive bonding. chemical bond forming Extrusion, injection molding, thermoforming Chemical vapor deposition, composite materials processing, food processing, glass technology, optical fiber drawing, powder metallurgy, sintering, sputtering, microgravity materials processing

turing processes, in which the heat transfer and fluid flow aspects are of crucial importance in determining the properties of the final product, are also often categorized as thermal-based manufacturing or materials processing. A few typical thermal manufacturing processes are sketched in Fig. 7.1, including optical glass fiber drawing, continuous casting, mold casting, and plastic screw extrusion. A fairly wide range of important manufacturing techniques involve thermal processing of materials. Consequently, a tremendous research effort has been directed at the thermal transport that is of interest in such processes. Several books are concerned with the area of manufacturing and materials processing. Most of these discuss the important practical considerations and manufacturing systems relevant to the various processes, without considering in detail the underlying thermal transport and fluid flow (Doyle et al., 1987; Schey, 1987; Kalpakjian, 1989). However, a few books have also been directed at the fundamental transport mechanisms in materials processing (Szekely, 1979; Ghosh and Mallik, 1986). Several other books consider specific manufacturing processes from a fundamental standpoint (Avitzur, 1968; Altan et al., 1971; Fenner, 1979; Easterling, 1983). In addition, there are several review articles and many symposia volumes on

260

Jaluria

Figure 7.1 Sketches of a few common thermal manufacturing processes: (a) glass fiber drawing; (b) continuous casting; (c) mold casting; (d) plastics screw extrusion,

thermal aspects of materials processing that have appeared in the recent years (Hughel and Bolling, 1971; Kuhn and Lawley, 1978; Chen et al., 1983; Viskanta, 1985, 1988; Li, 1985). It is clear that a substantial amount of work has been done on the heat transfer and fluid flow phenomena underlying materials processing. This chapter discusses the relevant funda-

Modeling of Manufacturing Systems

261

mental considerations and presents the numerical modeling of such manufacturing processes. Many important considerations arise when dealing with the numerical modeling of materials processing (Table 7.2). Most relevant processes are time dependent, since the material often undergoes a given thermal variation in order to attain desired characteristics. Therefore, the variation with time is important, even though steady-state situations are also of interest in a few cases. Sometimes, a transformation of the variables in the problem can convert a time-dependent problem to a steady one. Most manufacturing processes involve combined modes of heat transfer. Conjugate conditions arise due to the coupling between conduction in the solid material and convection in the fluid. Radiation is frequently important in these processes. The material properties are often strongly temperature dependent, giving rise to strong nonlinearity in the energy equation (Lee and Jaluria, 1996a, 1996b). Also, the material properties may depend on the shear rate, as is the case for polymeric materials, which are generally non-Newtonian (Fenner, 1979; Jaluria, 1996). Therefore, material properties affect the transport processes and are, in turn, affected by the transport. This aspect often leads to considerable complexity in the mathematical modeling and in the numerical simulation. The material undergoing the thermal transport process may be moving, as in hot rolling or extrusion, or the thermal source itself may be moving, as in laser cutting or welding. Additional mechanisms

Table 7.2 Some Important Considerations in Materials Processing 1. Coupling of transport with material characteristics: different materials, resulting material structure, properties, behavior 2. Variable material properties 3. Complex geometries 4. Complicated boundary conditions 5. Interaction between different mechanisms: surface tension, heat and mass transfer, non-Newtonian flow, chemical reactions, free surface, powder, and particle transport, phase change, microstructure conversion 6. Inverse problems 7. Different energy sources: laser, chemical, explosive, gas, fluid jet, heat 8. System optimization and control

Jaluria

262

such as surface tension effects and chemical reactions are important in several cases. Complex geometry and boundary conditions are commonly encountered. Frequently, an inverse problem is to be solved to obtain the conditions that result in a given temperature field. Finally, the process is obviously linked with the manufacturing system design and operation. All these considerations make numerical modeling of materials processing very involved and challenging. Special procedures and techniques are often needed to satisfactorily simulate the relevant boundary conditions and material property variations. However, the results obtained are important and interesting, since these are generally not available in the existing heat transfer and fluid mechanics literature. Numerical simulation results also provide appropriate inputs for the design and optimization of the relevant thermal system for the manufacturing process. Therefore, it is important to accurately model the process, mathematically and numerically. Let us first consider the basic conservation principles and the appropriate governing equations for these processes.

GOVERNING EQUATIONS

7.2 7.2.1

General Equations

The governing equations for heat transfer and fluid flow in materials processing are derived from the basic conservation principles for mass, momentum, and energy. For a pure viscous fluid, these equations may be written as

-DP +pv.v=o Dt

DT = v pc, Dt

(7.1)

(kVT)

DP + p@ + Q + /IT E

(7.3)

where D/Dt is the substantial or particle derivative, given in terms of the local derivatives in the flow field by D/Dt= a/& + V - V. Here, p is the fluid density, V the local velocity vector, F the body force acting per unit volume of the fluid, T the stress tensor, C , the specific heat at constant pressure, k the thermal conductivity, /I the coefficient of thermal expansion, p the dynamic viscosity of the fluid, the energy source per unit volume, p the local pressure,

8Z2

(8.2)

Passive Thermal Control of Electronic Equipment

Figure 8.3

309

(a) Geometry of enclosure; (b) internal details of electronic package.

z-momentum :

310

Joshi

energy (fluid):

a(ue)+-a(ve) a(we) - - - (pr1Ra)1’2(fi +E +E ) ax ay az ax2 a y 2 az2 +

(8.5)

energy (chip):

(a2e , a2e ke) -

ay2

ax2

a22

+-=O 1 R,H,

energy (substrate) :

(a2e

a2e

ax2 a Y Z

a2e) az2

(8.7)

The appropriate nondimensional parameters are Ra = gflQ12/avkf, Pr = v/a, U = u / U O , V = u / U 0 , W = w / U 0 , U. = (gflQ/kr)1’2, 8 = ( T - TJ/(Q/lkf), P = p / p U & X = x/l, Y = y/l, 2 = z/l, H , = h$, and H , = h$. It is noted that the normalization scales for the velocities and temperature involve Q, since it is the driving parameter for the transport. Since the normalized velocities and temperatures are functions of Q additionally through their dependence on Ra, the dimensional changes in the transport variables with Q need to be carefully interpreted. The boundary conditions at the enclosure walls are as follows:

ae -=o, u=o, V = O , w = o ax x=x,; e = o , U = O , V = O , W = O ae Y = 0, x,; -=o, U = O , v=o, w = o ay

x=o;

2 = 0, x,;

ae -=o, dZ

U=O,

v=o, w = o

where X L = Le). In addition, the boundary conditions at the interfaces between two different materials are Ri(”> axn

=Rj(%) ax,

; ei = e j ,

U = 0,

v = 0, w = o

where X , is the coordinate along the outward normal of the surface (i.e. X , = X if the surface is in the Y - 2 plane), i a n d j refer to the two different materials ( s for substrate, c for chip, and f for fluid), and R , = k,/k, and R , = k,/k,. These are implicitly satisfied in the present computations

Passive Thermal Control of Electronic Equipment

311

through the use of the harmonic mean formulation for interface diffusivities, as discussed later. For this investigation, X , = 5.1, H , = 0.08, and H , = 0.21 were used to allow comparison with experimental data. The Prandtl number was chosen as 25, with R , = 2360. This corresponds to the use of fluorinert liquid FC-75 (Product Manual, 1985) as the coolant with a silicon chip. The Rayleigh number was varied over a range from 103 and 109, assuming R, = 575, which corresponds to an alumina ceramic substrate. A 1-cm2 chip operated with a power level of 1.5 W in FC-75 corresponds to a Rayleigh number of approximately 109. To study the thermal spreading along the substrate, values of R, = 0.5,5, and 50 were also employed with Ra = 106. Numerical Method

The governing equations are discretized using a control-volume approach as described by Patankar (1980). This approach uses control volumes for velocities that are staggered with respect to those for temperature and pressure; a power-law scheme for the differencing of dependent variables; a harmonic mean formulation for the interface diffusivities ; and the SIMPLER algorithm for velocity-pressure coupling. The solution is obtained from an initial guess through an iterative scheme using a line-byline tridiagonal matrix algorithm. The conjugate conduction in the chip and substrate is handled numerically by solving the same full set of momentum and energy equations throughout the entire enclosure, but with a large value of viscosity specified for the solid regions. The numerical solution is assumed converged when the maximum temperature change during successive iterations is less than 0.00oO1 times the maximum temperature for that iteration, and when overall energy balances on the enclosure, the chip, and the fluid are obtained within 1%. The solution was obtained throughout the entire enclosure; no symmetry condition was imposed at 2 = XL/2. This prevented the forcing of a symmetric solution where one may not exist, as may be the case if the plume above the chip became unsteady. As it turned out, the solutions were indeed symmetric for all conditions examined. Most of the results presented here were obtained using a 23 x 22 x 22 or a 24 x 22 x 22 nonuniform grid. Several points were chosen near the hot and cold surfaces to capture the thermal boundary layers. These points were often adjusted slightly for various Rayleigh number ranges (i.e., more points closer to the wall for higher Rayleigh numbers). Below Ra = 105,the grid in the remainder of the flow was fairly uniform. At higher Rayleigh numbers, in which the boundary layer regions accounted for smaller portions of the enclosure, several intermediate grid points were located between the thermal boundary layer

Joshi

312 Velocities

Isotherms

2.0'1 0%

\ \

- - - c c / ,

s.o-tc*--

- -

l.O*lo-'5.0

4 .O

3.0

2.0

1.0

0.0

I+ - - 0.0

1.0

- _. 2.0

3.0 A

4.0

0.0

1.0

2.0

3.0

4.0

Figure 8.4 U-V velocity vectors and isotherms for the X-Y plane at Z = 2.55 for R, = 575: (a, b) Ra = 103; (c, d) Ra = 10'; (e, f ) Ra = 10'; ( g , h) detail near chip for Ra = 10'.

5.0

313

Passive Thermal Control of Electronic Equipment

Ve!ocities

Isotherms

1.O'I 0-'5.0

,- 0.0

-0.001 1

1.0

2.0

3.0

4.0

3.0

0.0

2.0

1.0

3.0

4.0

I so therms

Vel o ci ti c s

2.715

0.0

0.8

0.4

Figure 8.4

and the core of the flow, to capture the outer region of the momentum boundary layer. The grid spacing within the core of the enclosure was relatively coarse.

314

Joshi

Effect of Rayleigh Number (R, = 575)

To describe the three-dimensional nature of the flow and heat transfer, data for a few selected planes are shown in Fig. 8.4 for Ra = 103, 106,and 10’. Figure 8.4 shows the U-V velocity vectors and the isotherms for the X - Y plane at 2 = 2.55 (i.e., the symmetry plane). At Ra = 103 (Fig. 8.4a), the primary flow is in the X - Y plane and it is characterized by a single cell that is disturbed only slightly near the protruding chip. The isotherms at Ra = 103in Fig. 8.4b indicate that the relatively strong U velocities convect low-temperature fluid from the cold wall toward the hot wall in the lower portion of the cavity and high-temperature fluid from the hot wall toward the cold wall in the upper portion. Note that the chip is nearly isothermal, mainly as a result of the high R , . When Ra = 106,Fig. 8.4c, a cellular flow is also evident, but the bulk of the primary flow in the vertical direction is confined to boundary layers along the hot and cold surfaces rather than spreading throughout the enclosure, as seen at Ra = 103.These boundary layers thicken significantly near the lower and upper walls due to three-dimensional effects of the horizontal walls. The flows around the upper and lower corners of the chip lead to weak jets of fluid moving toward the substrate at the top of the chip and away from the substrate near the bottom of the chip. The effect of the flow on the isotherms in Fig. 8.4d is similar as that in Fig. 8.4b, but with the temperature gradients confined to thermal boundary layers along the hot and cold walls. The remainder of the core of the enclosure, between the boundary layers, can be characterized as well-stratified. At Ra = 109 in Fig. 8.4e, the vertical primary flow is confined to even thinner boundary layers near the hot and cold walls. The vertical velocities are much greater in the region near and above the chip compared to that along the substrate or the cold wall. In particular, the velocities in the plume from the top of the chip are much higher than the rest of the flow. There is still a weak cellular structure to the primary flow. Like the Ra = 106 case, most of the horizontal flow between the hot and cold boundary layers occurs in regions near the top and bottom walls, but with velocities less than 10% of the maximum vertical velocity. The isotherms in Fig. 8.4f reveal very steep gradients near the hot and cold surfaces and a well stratified core. The cross convection is negligible everywhere. Details of the flow near the chip (Fig. 8.4g, h) present a clearer view of the growing thermal boundary layer along the chip face. Velocity vectors and isotherms for the Y-2 plane at X = 0.185 (through the center of the chip) are shown in Fig. 8.5. These plots are composite drawings which take advantage of the symmetry about 2 = 2.55, with isotherms shown on the right and velocity vectors shown on the left of each plot. The I/ velocity component near the substrate is fairly constant in

Passive Thermal Control of Electronic Equipment

315

Isotherms

Velocities

4.0I

Isotherms

Velocities 1.O' 10 -

e.0-10-',

' L

* tni

,I J.0-

ad+

, I F ' 0.0

1.0

2.0

5.0

4.0

3.0

0.04'

0.0

1.0

#MI

I 1

2.0

3.0

4.0

b) Velocities

Isotherms

1.0=10-'-

( * I 4

1.0I I

o,o

1 I

I n 1 1

Y l

Figure 8.5 W-V velocity vectors and isotherms for the Z-Y plane at X = 0.185 for R, = 575: (a) Ra = 103; (b) Ra = 106; (c) Ra = 10'.

316

Joshi

the 2-direction, except in the vicinity of the sidewalls for Ra = 103 in Fig. 8.5a. The W-velocity component near the substrate is outward, toward the sidewalls, near the bottom of the enclosure and inward near the top of the enclosure. At Ra = 106 (Fig. 8.5b), the strength of the primary flow (in the Y direction) again remains fairly constant across the 2-direction, with a smaller region of diminishing strength near the side walls. The one exception to this is the region above the chip, where the flow is slightly stronger than that along the rest of the substrate. This is a result of the emergence of a plume from the top of the chip. Near the upper edge of the substrate directly above the chip, the W velocity component is outward rather than inward as in Fig. 8Sa, as this plume encounters the solid top surface. At Ra = 109 (Fig. 8 . 5 ~ the ) plume grows to dominate the primary flow. When it reaches the top wall, it spreads horizontally outward toward the side walls, as well as toward the cold wall, creating a small circular region of very strong secondary velocities. The isotherms in Fig. 8.5 reveal spreading of the heat away from the chip in a nonsymmetrical manner for all three Ra. Also seen is the formation of a thin thermal boundary layer type region around the chip with increasing Ra. The thermal stratification in the enclosure liquid below the chip becomes evident with increasing Ra. Figure 8.6 presents the fractions of the net generated power that are lost from the various chip surfaces, the maximum chip temperature, and the temperature of the substrate at 2 = 1.125 and Y = 2.55, as a function of the Rayleigh number. Because of the large value of the ratio of substrate-tofluid thermal conductivity, conduction through the back of the chip to the substrate accounted for most of total heat loss from the chip-over 90% at Ra = 103 decreasing to 76% at Ra = 109.The heat loss from the front face was the next largest and increased with Rayleigh number as a result of the increase in the heat transfer coefficients. The other four surfaces of the chip accounted for almost negligible heat loss at Ra = 103 (insert in Fig. 8.6a), with substantial relative increases with increasing Rayleigh number. The nondimensional chip temperature decreased by more than an order of magnitude from 0.101 for Ra = 103 to 0.00735 for Ra = 109. For comparison, the chip temperature for the conduction-only solution was 0.21. Figure 8.6b also shows how the reduction in substrate conduction at higher Rayleigh numbers led to a greater difference between the substrate and chip temperatures. Effect of Substrate Thermal Conductivity (Ra = 10')

The decrease in the thermal conductivity of the substrate has a profound effect on the surface heat loss fractions and the maximum chip temperatures

Passive Thermal Control of Electronic Equipment

317

a ) Heat loss 1.o

0.8

I -m+ Front ~~

0.6

0.4

0.2

* Top * Bo!!om * Sides

Substrate

b) Temperatures 1

-----------_-

Chip temperature: conduction only solution

.o 1

Figure 8.6 Effect of Rayleigh number on (a) fraction of heat loss from the hot surfaces (chip and substrate) and (b) maximum temperature of chip and temperature of substrate at Z = 1.125 and Y = 2.55.

318

Joshi

a) Heat loss 1.o

I Substrata

- 0

0.4

100

1000

b) Temperatures 0.1 0

0.08

0.06

0.04

0.02

0.00

t .1

- . - --.,, 1

. .

..--I

100

- . . . ..1000

U. Figure 8.7 Effect of substrate thermal conductivity on (a) fraction of heat loss from the hot surfaces (chip and substrate) and (b) maximum temperature of chip and temperature of substrate at Z = 1.125 and Y = 2.55.

(Fig. 8.7a). As R , is reduced, the fraction of loss through the substrate falls, while fractions from the various chip surfaces increase. At R , = 5, the heat loss from the front face is the largest, while the substrate fraction has

Passive Thermal Control of Electronic Equipment

319

dropped to only 14%. At R, = 0.5, the substrate loss is only 1%. The chip temperature increases substantially with decreasing R,. As seen in Fig. 8.6b, it increases from 0.02 for the baseline case to 0.06 for R, = 5. For R, < 5, further reductions of R , lead to only moderate increases in chip temperature, since substrate conduction plays a minor role. Comparison with Experiments

As mentioned, the geometry employed here was similar to one of the two configurations in the experiments performed by Joshi and Paje (1991). A key difference was that in the experiments the uniform heat generation was confined to the 1.52-mm square, 0.4-mm-thick silicon chip (or die as it is often called). This was located within a 20-pin “leadless chip carrier” package 8.9 mm square and protruding 1.9 mm from the substrate (Fig. 8.3b). The die was attached to the ceramic package through thin layers of gold and tungsten. An air space between the die and the lid of the package (made of kovar alloy) provided an additional internal resistance to heat transfer. The package itself was elevated slightly above the substrate, resting only on the 20 solder joints. Measurements of steady temperatures on the chip and selected substrate locations were reported in three dielectric liquids for a range of power levels. The 3-D code was run for this particular package design with a power level of 1.84 W with FC-75 as the coolant. A slightly modified form of Eq. (8.6) was used to prescribe the heat generation only within the chip. Using properties evaluated at the mean temperature between the lid and the cold wall (from the experiments), this corresponded to Ra = 1.15 x log and Pr = 24. Comparisons of measured substrate and chip temperatures with the computations are shown in Table 8.1. With the exception of the substrate temperature directly behind the package, the results are within - 9% and + 15%. Given the complicated design of the package, which could not be modeled in all its detail, and the uncertainty associated with the contact resistance between the package and the substrate, these results are satisfactory. The numerical simulation revealed very large temperature gradients along the substrate in the Y - and 2-directions directly behind the package, the region where a large discrepancy (35%) was observed between the numerical and experimental results. In the experiments, three factors could have led to locally increased spreading in this region, which might account for the lower measured temperatures. First, a high thermal conductivity paste was used to attach the thermocouple to the substrate and the size of this connection was sufficient to cover a region where the numerical results showed large temperature variations. Second, behind the substrate was a

Joshi

320 Table 8.1 Comparison of Numerical and Experimental Results for Lead less Chip Carrier Package Location Chip

8 experiment

8 numerical

0.00197

0.0226

0.0043 0.0072 0.0032 0.0033 0.0035

0.0048 0.0097 0.0029 0.0037 0.0039

Subst rate' 4. (Z = 2.55, Y = 3.98) 5. (Z = 2.55, Y = 2.55) 6. (Z= 2.55, Y = 1.12) 7. (Z = 1.12, Y = 3.95) 8. (Z= 1.12, Y = 2.55)

o/o

diff

+ 15 + 12 + 34 -9 + 12 +11

Numbers designate references to thermocouples used by Joshi and Paje ( 1991). a

slab of plexiglass, which might also have acted as an additional thermal path for spreading. In addition, heat loss from the edgeJ of the enclosure might have resulted in lower temperature throughout the entire enclosure. This may explain why most of the numerically predicted temperatures are above those measured experimentally. 8.4.2

Application of Solid-Liquid Phase-Change Materials for Passive Thermal Control of Electronic Packages

With increasing miniaturization of electronic products, there is a trend toward monochip packages with high power dissipation. For such packages, the use of phase-change materials (PCM) may provide thermal control for the entire duration of time the package is powered. These materials store thermal energy in the form of the latent heat of fusion, which can be rejected later to reverse the phase change. The PCMs to be used for thermal control purpose must meet several criteria. First, their melting point should be below the maximum allowable operating temperature range of the component. They should also be nonflammable, nonexplosive, and noncorrosive. For design purposes, their thermophysical properties must be available. Bentilla et al. (1966) identified four organic paraffins that meet these criteria. An experimental and analytical study of PCM for thermal control of electronics was performed by Witzman et al. (1983). Their analytical model uses correlations for phase-change heat transfer. Their experiment showed the potential of using PCM for cooling high-power modules for a substantial amount of time. Snyder (1991) performed a two-dimensional analysis of a PCM cooling scheme for an electronic module. Melting of PCM was assumed to be driven by conduction only, and effects of natural

321

Passive Thermal Control of Electronic Equipment Leads

Lead h

Encapsulant

Sides

Top -.

I[ Board

1 , \

\&E!!

\ Paddle

Figure 8.8 Details of the plastic quad flat package (dimensions in mm): (a) plan and (b) cross-sectional views.

convection in the melt pool were neglected. Ishizuka and f*ckuoka (1991) used a metallic eutectic layer of Bi/Pb/Sn/In with a melting point of 57°C under a simulated electronic package. Their experimental data show that the operating temperature rise of the component can be arrested for a significant amount of time. They also performed a one-dimensional network analysis of their PCM cooling technique. The computational studies of cooling of electronics in literature often consider simple geometries and boundary conditions. Heat transfer from discrete heat sources is often approximated by uniform heat flux from packages with internal details of the actual electronic package omitted. With these simplifications, the dominant heat transfer path from the package

Joshi

322

cannot be ascertained. Information about relative distribution of heat flow for these various paths is important for a PCM cooling system design to take advantage of the primary heat flow path from the package. In a recent study by Pal and Joshi (1996), a three-dimensional computational model was used to predict the performance of PCM cooling of a plastic quad flat package (PQFP). This study is next discussed in some detail. Computational Model

A 208-pin plastic quad flat package was simulated. Details of the package and materials were obtained from a study of the thermal model of a P Q F P by Rosten and Viswanath (1994). The dimensions of the package are shown in Fig. 8.8. Several simplifications were made to model the package. The lead frame was assumed to be in the same plane as the silicon chip. In order to reduce the number of control volumes, every group of 5.2 leads of the package were lumped as one equivalent lead, reducing the total lead count from 208 to 40. The lead frame in the plastic was treated as a mixture with uniform thermophysical properties. The solder joints for each lead were neglected. The computational domain considered is shown in Fig. 8.9, The package is mounted on 1-mm-thick ceramic or FR-4 printed wiring board (PWB) oriented in vertical direction. This assembly is placed in an enclosure of size 50 x 50 x 50 mm. A layer of PCM 10 mm thick is used under

i. /

Figure 8.9

PQFP Computational domain (dimensions in mm).

Passive Thermal Control of Electronic Equipment

323

the board. The reason for implementing the PCM under the board is that the primary heat flow path from the chip to the outside of the package is found to be through the leads by conduction. The right wall of the enclosure is assumed to be at a constant temperature of 25"C, and all other walls are adiabatic. Governing Equations

The governing conservation equations for mass, momentum, and energy can be written as continuity :

x-momen t um :

- (pu) + - (puu) + -

= -(P

g)+;

$) +;

(PW4

E) 2 + s, -

(8.9)

y-momentum :

= -(P

g) + a ay

">+ 2az E) 0+ s,, (p

(8.10)

z-momen tum :

(8.11)

energy :

(8.12)

The same set of equations applies for various materials. The expressions for coefficients and source terms for various materials in these equations are

324

Joshi

Table 8.2

Source Terms for Various Materials

Pgm

Air PCM

C(1 ~ ~

U+

- To)

C(1 c3 + b

0 C(1 ~ ~

0 w+

+ P g N - T,) Silicon

Other solids

listed in Table 8.2. The thermophysical properties of various materials are provided in Table 8.3. The heat dissipation within the silicon chip was assumed to be volumetrically uniform. To handle conjugate conduction in the solids, the same single-domain approach described in the previous section was used. Modeling of Phase Change

The phase change was handled using a single-domain, enthalpy porosity technique (Brent et al., 1988). In this method, the absorption of latent heat Table 8.3

Thermophysical Properties Used for Computation

Materials

Thermal conductivity (W/m°C)

Ai r 0.0261 FR-4 0.35 Ceramic 18.00 (C-786) Leads and 385 paddle (copper) 0.31 Encapsulant Lead frame 154.17 encapsulant mixture Chip 154 86(300/T)4’3 (silicon) PCM 0.23 (n-eicosene) Melting point of PCM: 37°C

Specific heat (J/kS-K)

Dynamic viscosity (kg-m/s)

1005 1600 840

1.85 x 10-6 1030 1030

8933

385

1030

1070 4215.2

100 214

1030 1030

69 1

2330

1030

795

2050

3.57 x 10-3

Density (kg/m3)

1.177 1938 3875

Latent heat per unit mass: 241 kJ/kg

Passive Thermal Control of Electronic Equipment

325

during melting is included as a source term S, in the energy equation (Table 8.2). Latent heat content of each control volume in the PCM is evaluated after each energy equation iteration cycle. Based on the latent heat content, an effective porosity ( E = A H / L ) for each control volume is determined. Control volumes that contain molten PCM have E = 1 and control volumes containing solid PCM have E = 0. The control volumes with values of E between 0 and 1 are treated as mushy. Even though the phase change is assumed to be isothermal, the idea of mushy zone is introduced to gradually “switch off” the velocities from liquid to solid at the interface. The “switching off” is controlled by the source terms S,, S,, and S, in the momentum equations (Table 8.2). The solid-liquid interface is assumed to correspond to the E = 0.5 line. Numerical Procedure and Validation

Grid sizes were selected carefully to resolve the internal details of the package. Local grid refinement was required to model the chip, paddle, lead frame, and leads of the package. The grid outside the package was relatively coarse. Fine grid resolution in air was used near all solid walls to capture the boundary layer effects. Based on grid size testing, a 34 x 39 x 39 grid was selected for all computations. Though the domain is symmetric in the z-direction about the midplane, the analysis was performed for the entire enclosure due to the fact that at higher Rayleigh numbers the flow structure in the enclosure may be time dependent with no symmetry along the midplane. A finite-volume-based numerical algorithm, SIMPLER (Patankar, 1980) is used to solve the governing conservation equations. This code is fully implicit in time for transient computations. However, a control of time step was necessary to achieve convergence due to the complex and transient heat transfer stages involved. During the initial phase, when conduction is the dominant mode of heat transfer, a time step of 10 s was used. However, after the natural convection begins to dominate, the time step had to be reduced to 5 s. A further reduction to 2 s was necessary when melting was in progress. An optimum set of relaxation factors for best convergence was obtained after trial and error. Convergence for a time step is assumed when the sum of normalized residuals for temperature was less than 1 x 10-4. Simultaneously, a global energy balance residual of less than 1 x 1 O W 3 is prescribed for convergence. Results for Computations with FR-4 Board

Computations were performed for six cases. These cases consider two different board materials and two different power levels to estimate their effects

Joshi

326 Table 8.4

Summary of Parameters for Various Cases

Case

Power (watts)

PWB material

Computation

A B C D E F

1 1 1 1 3 3

FR-4 FR-4 Ceramic Ceramic Ceramic Ceramic

With PCM Without PCM With PCM Without PCM With PCM Without PCM

on the performance of PCM cooling. Board materials were FR-4 and ceramic. A summary of various cases is presented in Table 8.4. Initially the computational domain was assumed to be at uniform temperature of 25°C. Figure 8.10 shows the timewise variation of maximum temperature of the chip and the temperature at the geometric center of the board for cases A and B. Three different stages of heat transfer can be observed for each curve. During the initial stage, a rapid temperature rise is seen. During this period, the heat transfer is largely by conduction from the chip to the lead frame, plastic encapsulating material, and surrounding air. The resulting rate of temperature rise for the board is slower than that of the chip. The second stage is characterized by a slowdown in the rate of temperature rise of the chip and the development of natural convection in the air. The air heated by the package and the PWB circulates and contacts the isothermal right wall. Heat is transferred from the hot air to the wall in the boundary layer adjacent to it. The third phase for case A begins when the heat conducted across the PWB raises the PWB/PCM interface temperature to the PCM melting point. During this phase, the PCM melts isothermally. At the initial times of the third stage, the melting is driven by conduction heat transfer from the board. This is characterized by a thin planar melt layer as shown in Fig. 8.10b, d. However, as melting progresses, natural convection begins to develop in the melt pool, due to which the solid liquid interface changes as seen in Figure S.lOf, h. Due to the lower thermal conductivity of the FR-4 board, heat spreading along the y- and z-directions are relatively weaker than in x-direction, due to which the melting process is localized around the footprint area of the package and is not spread out uniformly along the board. After 1 h of operation, the melting process is still in progress and natural convection is dominant in the melt pool (Fig. 8.10g, h). During this phase, the maximum temperature of the package is stabilized at 61°C. For the case without PCM (case B), the third stage is a mere continuation of the second stage, with the package and board approaching a steady-state condition. However, it was

Passive Thermal Control of Electronic Equipment

327

Figure 8.10 Isotherms, velocity vectors, and melt shapes for 1 W of power and FR-4 as the PWB material: (a) isotherms at t = 400 s, ( b ) melt shape at t = 400 s, (c) isotherms at t = 800 s, (d) melt shape at f = 800 s, (e) isotherms at t = 2200 s, and (f) melt shape at t = 2200 s.

Fig. continues

Joshi

328

Figure 8.10 Continued. (9) Isotherms at t = 3600 s, (h) melt shape at and (i) velocity vectors at t = 3600 s.

3600

found that after 1 h the package had not reached a steady-state condition. One run was performed with very large time step to assess the steady-state temperatures, and the maximum chip temperature was found to be 136°C. Figure 8.101 shows the velocity vector plot at the ( z = z1/2) plane and indicates a natural convection flow in the enclosure. The velocity field shows a strong plume generated above the package which moves up and impinges on the upper surface and creates a circulation cell throughout the enclosure. 8.5

SUMMARY AND FUTURE DIRECTIONS

Traditional semiempirical techniques such as thermal network methodology or modified conduction analysis require input information in the

Passive Thermal Control of Electronic Equipment

329

form of convection coefficients. For most electronic cooling applications, applicable correlations do not exist. Conjugate simulations using the CFD/CHT approach do not require a priori specification of the convection coeficients at the component faces. This chapter has focused on the application of these techniques for the analysis of passively cooled electronic equipment. With the push toward reduced times to market of modern electronic products, the need for such analyses is likely to grow considerably. While these techniques offer unprecedented potential capability for thermal characterization at the various levels of packaging, a number of issues need to be addressed to make them more useful for proactive design of electronic products. Currently, one of the most significant limitations of the CFD/CHT approach is the large computational effort involved in simulating a realistic electronic product. In order to accurately predict thermal performance of the complete electronic system, modeling at all three levels of packaging (i.e., component, board, and system) must be performed in an integrated and efficient manner. Innovative approaches for carrying out such global/local analyses are required. Also, very little experimentally validated information exists on the handling of transitional and turbulent transport in electronic cooling applications. The importance of thermal radiation in air-cooled systems and its incorporation in a general conjugate analysis needs more attention. Finally, one of the impediments in achieving better predictive capability with such techniques is the unavailability of thermophysical properties data for various electronic packaging materials. In many instances, for example for epoxy-fiberglass circuit boards, strong anisotropic behavior is found for the thermal conductivity. In-plane and out-of-plane values can be different by as much as a factor of 3. More research is needed for coming up with techniques to determine effective properties that can provide adequate accuracy in temperature predictions for such cases. ACKNOWLEDGMENTS

The author acknowledges the assistance of Dr. D. Pal in the preparation of this chapter. NOMENCLATURE

b c, C

constant in the porosity source term specific heat at constant pressure [J/kg-K] morphological constant

Joshi

330

gravitational acceleration [m/s2] heat transfer coefficient [W/m2-K] thickness of protrusion [m] h C thickness of substrate [m] hs AH latent heat component of enthalpy [kJ/kg] k thermal conductivity [W/m-K J 1 protrusion length [m] L latent heat of fusion [J/kg-K] L e enclosure length [m] Nu 4,.L/k,-(T- q),Nusselt number pressure [N/m2] P Pr v/a, Prandtl number of fluid heat flux crossing a solid surface [W/m2] 4i Q heat generation rate [W] Q volumetric heat generation rate [W/m3] Ra gbQ12/vak,, Rayleigh number Rc kc/k,, ratio of protrusion thermal conductivity to fluid thermal conductivi t y k J k , , ratio of substrate thermal conductivity to fluid thermal conductivity source term [W/m3] S T temperature [K J T temperature at the cold enclosure wall parallel to the substrate [K] U velocity component in x-direction [m/s] velocity component in y-direction [m/s] U w velocity component in z-direction [m/s] 9

)'I

Greek Symbols a

8 p

v p

fluid thermal diffusivity [m2/s] coefficient of volumetric expansion [l/K] ( T - T , ) / ( Q / k , I) nondimensional temperature dynamic viscosity [kg-m/s] kinematic viscosity [m2/s] fluid density [kg/m3]

Subscripts

0 a c

ambient air protrusion (chip)

Passive Thermal Control of Electronic Equipment

f H i

rn p s

33 1

fluid enthalpy condition at a solid surface melting point phase change material substrate

REFERENCES Bar-Cohen A. Thermal management of electronic components with dielectric liquids. Proceedings of ASME/JSME Thermal Engineering Joint Conference. Vol 2. pp xv-xxxix. Bergles AE, Bar-Cohen A. Direct liquid cooling of microelectronic components. In : Bar-Cohen A, Kraus AD, eds. Advances in Thermal Modeling of Electronic Components and Systems. Vol 2. ASME Press, 1990, pp 233-342. Bentilla EW, Sterrett KF, Karre LE. Research and development study on thermal control by use of fusible materials. Northrop Space Laboratories, Contract No. NAS 8-1 1163, NASA Document No. N66-26691, 1966. Brent AD, Voller VR, Reid KJ. Enthalpy-porosity technique for modeling convection-diffusion phase change: application to the melting of a pure metal. Numer Heat Transfer 13: 297, 1988. Chan AMC, Banerjee S. Three-dimensional numerical analysis of transient natural convection in rectangular enclosures. ASME J Heat Transfer 101 : 114, 1979. Heindel TJ, Ramadhyani S, Incropera FP. Laminar natural convection in a discretely heated cavity I. Assessment of three-dimensional effects. ASME J Heat Transfer 117: 902, 1995a. Heindel TJ, Ramadhyani S, Incropera FP. Laminar natural convection in a discretely heated cavity 11. Comparisons of experimental and theoretical results. ASME J Heat Transfer 117 910,1995b. Ishizuka M, f*ckuoka Y. Development of a new high density package cooling technique using low melting point alloys. Proceedings ASME/JSME Thermal Engineering Joint Conference. Vol2, 1991, pp 375-380. Joshi Y, Kelleher MD, Benedict TJ. Natural convection immersion cooling of an array of simulated electronic components in an enclosure filled with dielectric fluid. In Bergles AE, ed. Heat Transfer in Electronic and Microelectronic Equipment. Hemisphere, 1990, pp 445-468. Joshi Y, Paje RA. Natural convection cooling of a ceramic substrate mounted leadless chip carrier in dielectric liquids. Int Comm Heat Mass Transfer 18: 39, 1991. Joshi Y, Willson T, Hazard SJ 111. An experimental study of natural convection from an array of heated protrusions on a vertical surface in water. ASME J Electron Pack 1 1 1 : 121,1989a. Joshi Y, Willson T, Hazard SJ 111. An experimental study of natural convection

332

Jos hi

cooling of an array of heated protrusions in a vertical channel in water. ASME J Electron Pack 11 1 : 33, 1989b. Kelleher MD, Knock RH, Yang KT. Laminar natural convection in a rectangular enclosure due to a heated protrusion on one vertical. Part I. Experimental investigation. Proceedings of the 2nd ASM E-JSME Joint Thermal Engineering Conference. Vol 11. Honolulu Hawaii, 1987, pp 169-178. Keyhani M, Chen L, Pitts DR. The aspect ratio effect on natural convection in an enclosure with protruding heat sources. Presented at the AIAA/ASME Thermophysics and Heat Transfer Conference, Seattle, WA, 1990. Keyhani M, Prasad V, Cox R. An experimental study of natural convection in a vertical cavity with discrete heat sources. ASME J Heat Transfer 110: 616. 1988. Kuhn D, Oosthuizen PH. Three-dimensional natural convective flow in a rectangular enclosure with localized heating. Proceedings of AIAA Thermophysics Conference, 1987, pp 55-62. Lankhorst AM, Hoogendoorn CJ. Three-dimensional calculations of high Rayleigh natural convection flows in enclosed cavities. Proceedings of the National Heat Transfer Conference, Houston, TX. Vol3. 1988, pp 463-470. Lee JJ, Liu KV, Yang KT, Kelleher MD. Laminar natural convection in a rectangular enclosure due to a heated protrusion on one vertical wall Part 11. Numerical simulations. Proceedings of the 2nd ASME-JSME Joint Thermal Engineering Conference, Vol 11. Honolulu, Hawaii, 1987, pp 179-1 85. Liu KV, Yang KT, Kelleher MD. Three-dimensional natural convection cooling of an array of heated protrusions in an enclosure filled with a dielectric fluid. Proceedings of the International Symposium on Cooling Technology for Electronic Equipment. Honolulu, Hawaii, 1987a, pp 486-497. Liu KV, Yang KT, Wu YW, Kelleher MD. Local oscillatory surface temperature responses in immersion cooling of a chip array by natural convection in an enclosure. Proceedings of the Symposium on Heat and Mass Transfer in Honor of B. T. Chao. University of Illinois, Urbana-Champaign IL, 1987b. Mallinson GD, de Vahl Davis G. Three-dimensional natural convection in a box: a numerical study. J Fluid Mech 83: 1, 1977. Pal D, Joshi Y. Application of phase change materials for passive thermal control of plastic quad flat packages (PQFP): A computational study. Num Heat Transfer, Part A 30: 19, 1996. Park KA, Bergles AE. Natural convection heat transfer characteristics of simulated microelectronic chips. ASME J Heat Transfer 109: 90, 1987. Patankar SV. Numerical Heat Transfer and Fluid Flow. New York: Hemisphere McGraw Hill, 1980. Prasad V, Keyhani M, Shen R. Free convection in a discretely heated vertical enclosure: effects of Prandtl number and cavity size. ASME J Electron Pack 112: 63, 1990. Product Manual Fluorinert Liquids 3M Corporation, Minneapolis, MN, 1985. Rosten H, Viswanath R. Thermal modelling of the Pentium processor package. Proceedings of the Electronic Component and Technology Conference 1994.

Passive Thermal Control of Electronic Equipment

333

Sathe S, Joshi Y. Natural convection arising from a heat generating substratemounted protrusion in a liquid-filled two-dimensional enclosure. Int J Heat Mass Transfer 34 : 2 149, 1991. Semiconductor Industries Association Roadmap, 1994. Snyder KW. An investigation of using a phase-change material to improve the heat transfer in a small electronic module for an airborne radar application. Proceedings of the International Electronics Packaging Conference. Vol 1. 1991, pp 276. Viskanta R, Kim DM, Gau C. Three-dimensional numerical natural convection heat transfer of a liquid metal in a cavity. Int J Heat Mass Transfer 29: 475, 1986. Witzman S, sh*tzer A, Zvirin Y. Simplified calculation procedure of a latent heat reservoir for stabilizing the temperature of electronic devices. Proceedings of the Winter Annual Meeting of the ASME. HTD Vol28. 1983, pp 29-34. Wroblewski DE, Joshi Y. Computations of liquid immersion cooling for a protruding heat source in a cubical enclosure. Int J Heat Mass Transfer 36: 1201, 1993. Yang HQ, Yang KT, Lloyd JR. Flow transition in laminar buoyant flow in a threedimensional tilted rectangular enclosure. Proceedings of the 8th International Heat Transfer Conference, San Francisco, CA, 1986, pp 1495-1 500.

This page intentionally left blank

Computational Fluid Dynamic Techniques in Air Quality Modeling M CN C-Environmental

Mehmet T Odman* Programs. Research Triangle Park. North Carolina

Armistead G Russell Georgia Institute of Technology. Atlanta. Georgia

9.1 Background and Introduction . . . . . . . . . . 9.1.1 Eulerian Models . . . . . . . . . . . . . 9.2 Advection Schemes . . . . . . . . . . . . . . 9.2.1 Classification of Schemes . . . . . . . . . . 9.2.2 Brief Description of Schemes . . . . . . . . . 9.3 Evaluation of Advection Schemes . . . . . . . . . 9.3.1 Description of Tests and Performance Measures Used 9.3.2 Evaluation Cases . . . . . . . . . . . . . 9.3.3 Results and Discussion . . . . . . . . . . . 9.4 Evaluation of Advection Schemes in an Air Quality Model and the Impact on Predictions . . . . . . . . 9.4.1 Tracer Experiments . . . . . . . . . . . . 9.4.2 Evaluation of Advection Schemes with Full Chemistry 9.5 Summary . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . .

. . . . . . .

336 340

. . . . .

389 390 391 396 397

345 345 350 363 363 . . . . . 367 . . . . . 387

*Current affiliation: Georgia Institute of Technology. Atlanta. Georgia .

335

336

9.1

Odman and Russell

BACKGROUND AND INTRODUCTION

About one fourth of the population in the United States lives in areas which experience episodes of photochemical air pollution, and in 1990 almost 100 cities were out of compliance with the National Ambient Air Quality Standard (NAAQS) (NRC, 1992). Air pollution abatement would reduce the risks of adverse human health effects and vegetation damage and may slow the rate of material degradation as well. However, it is desirable to achieve these environmental standards in a cost-effective manner. Advanced computer models follow pollutant species emitted from anthropogenic and biogenic sources as they are transported downwind to receptor areas and simulate their complex chemical interactions with other species in the atmosphere. Currently these models represent the most scientifically sound foundation for testing alternative control strategies; thus, they are increasingly being used as the basis for regulations. Billions of dollars may be spent to comply with these regulations, so it is important to understand the uncertainty in the underlying models. While many (if not most) of the uncertainty in model predictions comes from uncertainty in the model inputs, there is some introduced by the model itself, e.g., by the numerical advection routines employed. One of the major problems facing our society that has grown out of industrialization is the deterioration of air quality. In particular, problems such as acid deposition, smog, global climate warming, and stratospheric ozone depletion pose significant threats to both human health and welfare and related ecological damage. There are a number of facets to these problems, and the question is how do we address and mitigate them? Like most environmental problems, they are very complex. However, there is a tremendous payoff for identifying effective solutions. That is the role of air quality models. Here, we discuss the types of numerical techniques used by air quality models, particularly the more advanced models that follow atmospheric transport and chemistry, e.g., those describing urban and regional smog, acid deposition, global atmospheric chemistry, etc. While operating at different scales, and often being used for different problems, they have very similar characteristics. First, the problems are described by the same set of conservation equations. Second, the numerical techniques employed are similar to those used in more traditional computational fluid dynamics applications. As an introduction, it is instructive to look at how we currently manage air pollution (Fig. 9.1). It is a feedback process, examining whether our current air quality meets the desired goals, identifying candidate control strategies (e.g., different ways to decrease emissions of air pollutants and their precursors), and testing how well those controls work. This last step is

Air Quality Modeling

337

Figure 9.1 Schematic of the air quality control planning process Air quality models are central to the process Models make it possible to determine how emissions affect pollutant concentrations and the resulting environmental effects

the determination of source-air quality relationships, and is generally done using an air quality model. To do an adequate job in this step, the model used must contain the applicable processes important to the problem at hand, and numerical routines that are commensurable in terms of accuracy and speed. For example, for urban smog, the model must treat both the chemistry and physics of the atmosphere. One of the more important processes is advection, and the models used to numerically follow the transport of species are very similar to those used in more traditional computational fluid dynamic (CFD) applications. However, the dominance of some of the processes can be quite different. Before discussing the different numerical methods, it is instructive to develop the set of equations being solved. Atmospheric pollutant dynamics are mathematically described by the species conservation equation :

i = 1, 2, 3, ..., n

(9.1)

where ci is the concentration of species i, U is the wind velocity vector, D i is the molecular diffusivity of species i, Ri is the net production (depletion if

Odman and Russell

338

negative) of species i by chemical reaction, Si (x, t ) is the emission rate of i at location x, p is the air density; and n is the number of species. R can also be a function of the meteorological variables (e.g., temperature, T). In essence, this equation states that the time rate of change of a pollutant (term 1) depends on convective transport, diffusion, and chemical reaction of that pollutant. The result is a set of n coupled, nonlinear partial differential equations. The coupling and the nonlinearity are introduced by the reaction term. This equation is subject to the initial conditions on the species concentrations as discussed below. There are known analytical solutions for this equation for only rather simple wind fields, diffusivity profiles, and chemical interactions. Thus, solution is usually accomplished by computational methods. In actual practice, the situation is further complicated by the turbulent nature of the atmosphere, making the description (and our knowledge) of the wind velocities more problematic. For one, wind velocity fluctuations chaotically occur at time and spatial scales much smaller than we can computationally afford to follow over a typical domain of interest. Luckily, we usually are not interested in the concentration fluctuations at the very small scales. As usual, the impact of turbulence is included by the use of Reynolds’ decomposition, where the velocity components and concentrations are split into an average component ((U) and (c)) and a fluctuating part (U’,c’), e.g., U = (U)

+ U’

(9.2)

Different averaging procedures are discussed in Seinfeld (1985). The result of this process is an equation for the dynamics of the averaged concentration component:

In practice, the turbulent diffusion term, (u’c’), is much larger than the molecular diffusion term, so the latter is usually neglected. However, calculating the turbulent diffusion is a challenge due to the closure problem. Thus, it is usually parameterized by the introduction of an effective turbulent diffusivity tensor, K j i :

This is analogous to Fickian diffusion, and is called K-theory. The turbulent diffusivity is not a fluid property and is a function of the fluid dynamics. Use of this formulation introduces limitations into the applicabil-

Air Quality Modeling

339

ity of the resulting model, primarily in a lower limit of the spatial scale. However, this is generally not a problem for regional models. The resulting equation upon which most air quality models are based is

+ (S&,

t)),

i = 1, 2, 3, ..., n

(9.5)

and is often referred to as the atmospheric diffusion equation, or ADE (Seinfeld, 1985). Air quality models, based on the ADE, are applied to follow pollutant dynamics over a designated spatial domain for specific periods of time. Examination of Eq. (9.5) shows that if there are no chemical reactions ( R = 0), or if R is linear in ( c i ) and uncoupled, then a set of linear, uncoupled differential equations are formed for determining pollutant concentrations. This is the basis of transport only and transport with linear chemistry models (which, for brevity, will be called transport models). Transport models are used for studying the effects of sources of CO and primary particulate matter on air quality, and have been used for sulfate formation and deposition, but they are not suitable for studying reactive pollutants such as 0, , NO,, HNO, ,and secondary organic species. The solution of the ADE requires specification of initial and boundary conditions. The initial conditions generally used are those specifying the concentration at the beginning of a simulation:

Horizontal boundary conditions are usually taken as those concentrations at the modeling domain boundary: c(Q t ) = p " n d * r y (0

(9.7)

where Cl is the boundary of the modeling domain. The ground-level boundary condition is a statement of the balance between deposition, vertical diffusion, and ground-level emissions :

The upper-level boundary condition can take two forms. First is a specification of the species concentrations at the top of the domain, similar to the specification of the horizontal boundary conditions. However, this assumes knowledge of the atmospheric concentrations well above the ground which are often not known. An alternative boundary condition is to assume a

Odman and Russet1

340

negligible concentration gradient :

9= o dZ

(9.9)

This, in essence, assumes that the best estimate of the species concentration (actually, the mixing ratio) above the modeling domain is the prediction just below the top. Use of the former upper-level condition (i.e. specifying the concentration) can lead to significant, artificial input (or loss) of pollutant mass by diffusion if the specified concentration is too high (or too low).

9.1.1 Eulerian Models

There are two distinct reference frames from which to view pollutant dynamics. The most natural is the Eulerian coordinate system which is fixed at the earth’s surface. In that case, a succession of different air parcels are viewed as being carried by the wind past a stationary observer. The second is the Lagrangian reference frame which moves with the flow of air, in effect maintaining the observer in contact with the same air parcel over extended periods of time. Because pollutants are carried by the wind, it is often convenient to follow pollutant evolution in a Lagrangian reference frame, and this perspective forms the basis of Lagrangian trajectory. Both Eulerian and Lagrangian models have been used to understand pollutant dynamics, though Lagrangian models have very severe limitations in their formulation. Given those limitations (they do not capture wind shear, which is important in the atmosphere), and that those models d o not use CFDtype techniques, we restrict the discussion below to Eulerian models. Eulerian “grid” models are the most complex, but potentially the most powerful, air quality models, involving the least restrictive assumptions. They are also the most computationally intensive. Grid models solve a finite approximation to Eq. (9.9, including temporal and spatial variation of the meterological parameters, emission sources, and surface characteristics. Grid models divide the modeling region into a large number of cells, horizontally and vertically (Fig. 9.2), which interact with each other by simulating diffusion, advection, and sedimentation (for particles) of pollutant species. Most of the current regional photochemical models are Eulerian models. Input data requirements for grid models include spatially and temporally resolved emissions (by species), meteorology (e.g., wind velocities, temperatures, solar insolation, etc.), topographic features, initial and background pollutant concentrations (for initial and boundary conditions), and domain definition. Eulerian grid models predict pollutant

Air Quality Modeling

341

T 7

Emissions 1. Anthropogenic -#* 2. Geogenic 3. Biogenic

Sources

Air Quality

Sources Pollutants

Numerical Solution Techniques

Concentrations Computed

* I

Meteorology

Transport (U,K,V,)

Land use

Geographical Features Figure 9.2 Schematic diagram of a typical air quality model, showing the model components and interactions.

concentrations throughout the entire modeling domain, usually including the airshed of interest and surrounding areas. Over successive time periods the evolution of pollutant concentrations and how they are affected by transport and chemical reaction can be tracked. Most air quality models use the time-splitting approach (Yanenko, 1971; Marchuk, 1975) for the time integration of this equation. In this approach, the time-marching operator of ADE is decomposed into discrete operators for horizontal transport, vertical mixing, chemical transformations, and source and removal processes. Although the method of time

342

Odman and Russell

splitting differs between models one such way is (9.10)

where

(9.11)

In essence, this means that at each time step the concentration field is first transported (diffused and advected) in the x-direction using the velocity field at time t. Next, the new concentration field (i.e., the one due to xtransport) is transported in the y-direction to get a new field. Next, the same is done in the z-direction. At this point the concentration field that existed at time t has been transported over the time step. Finally, the impact of chemistry over that time step is simulated. This gives the new field after one time step. Thus, the sequence of integrations is conducted as follows :

+ At) = Lx(c(t)) c**(t + Ar) = L,(c*(t)) c***(c + At) = L,(c**(t)) c(t + At) = L,(c***(t)) c*(t

where c*, c**, and c*** are intermediate calculations. An appropriate numerical technique (e.g., a one-dimensional advection-diff usion technique or chemical kinetics solver) is used for integrating each operator over the chosen time step. Next, the process is repeated but in reverse order (to reduce biasing and also to reduce the amount of time spent in solving the chemical kinetics, which accounts for most of the computational time), as shown in Eq. (9.10).This gives the new concentration field after evolving for two time steps. Thus, splitting is between horizontal transport along the two principal axes, the vertical direction, and a chemistry/source operator. While horizontal transport is more advection dominated, diffusive transport usually dominates vertically. Solution of the chemistry and source

Air Quality Modeling

343

operator uses much different techniques that are not discussed here. Of importance, though, is that the chemistry is the most intensive aspect of the solution, requiring about 85% of the computer time. Thus, it does not add significantly to the overall time to use a more accurate advection scheme. Further, the advection schemes, at present, appear to add the most error to the solution. As stated, horizontal transport in the atmosphere is advection dominated and can be represented by the horizontal advection equation (9.12)

where U and U are the advective velocities in the x- and y-directions, respectively. Equation (9.12) is known as the “flux form” of the advection equation. The same equation can be written in “advective form” as 84

- + U - + u - = o

(9.13)

where q is the mass mixing ratio related to the concentration c and density of air p as q = c/p. In the last two decades, the problem of advecting scalar fields (e.g., pollutant concentrations) has received significant attention, especially within the disciplines of computational fluid dynamics and meteorology. In a review article, Rood (1987) listed over 100 advection schemes for numerical solution of the advection equation. Several new schemes have since emerged; however, only a few are specifically designed for use in Eulerian air quality models. Many models split the horizontal advection equation and solve two one-dimensional equations, one in each direction, using the solution of one as the initial condition of the other:

ac a(uc) -+--=() at

(9.14)

ac a(0C) -+-=o at

These schemes will be referred to as 1-D. Others that solve the twodimensional form directly will be referred to as 2-D. Although using 1-D schemes is very common, it has been found that problems can arise due to this additional splitting (Flatoy, 1993; Odman and Russell, 1993). For example, even if the flow field is divergence free (i.e., V U = 0), this does not mean that &/dx and &/ay are both equal to zero. This automatically introduces an error because the divergence (or convergence) in one direction affects the intermediate solution, which is used as the initial condition a

Odman and Russell

344

to the other direction, and the convergence (or divergence) in the other direction may not yield the desired two-dimensional solution. In this regard, 2-D schemes may be more desirable. All the schemes discussed here are 1-D schemes with the exception of the semi-Lagrangian scheme, which solves the advective form in Eq. (9.13). Two-dimensional schemes that solve the flux form are usually more difficult to implement and computationally less efficient; very few have been tested in air quality models. For an example of a 2-D scheme solving the flux form the reader is referred to Odman and Russell (1991b). The 1-D schemes discussed here can also be used for the solution of vertical advection; however, since vertical transport in the atmosphere is diffusion dominated, selection of solver for vertical advection is not as critical an issue as it is for horizontal advection. There are several properties that an advection scheme must have in order to be useful in air quality models. As with all numerical methods, the numerical scheme for solving the advection equation must meet the convergence condition and correctly model the conservative, transportive, dissipative, and dispersive properties of the governing equation. A scheme is said to be convergent if the numerical solution approaches the true solution of the partial differential equation as the grid spacing and time-step size approach zero. Thus, if a scheme is convergent, one can obtain a numerical solution of any desired accuracy by reducing the grid spacing and the timestep size. For linear equations, consistency and stability are necessary and sufficient conditions for convergence (Lax’s equivalence theorem). In practice, the computational resources available limit the grid spacing and the time-step size. Therefore, numerical errors associated with using limited grid spacing and time-step sizes must be of concern. Due to the presence of pollutant sources and highly nonlinear chemical interactions between the concentration fields of different species, it is necessary to consider schemes with special properties : mass conservation, small numerical diffusion, and small phase errors. Advection schemes should be free of mass conservation errors to account accurately for pollutant sources and sinks and should have small numerical diffusion, since diffusion spreads a disturbance in every direction and smooths spatial gradients. They should also have small phase errors since disturbances that propagate at different speeds produce spurious oscillations such as “wiggles” or “ripples” in the numerical solution. These oscillations may lead to physically unrealistic negative concentrations, which are unacceptable in the presence of nonlinear chemistry in air quality models. Another source of spurious oscillations is the Gibbs phenomenon, which arises near steep gradients such as pollutant puffs or plumes due to the truncation of Fourier components of the solution. The nature of spurious oscillations leads to important concepts of “positive definiteness” and monotonicity in applying various schemes to “

”

Air Quality Modeling

345

air quality modeling. Schemes that do not allow negative concentrations are positive definite; however, these schemes may still allow oscillations such as overshoots in the solution. In general, monotonic schemes suppress all spurious oscillations. Since photochemical modeling starts with positive initial conditions, a monotonic scheme that does not create new extrema will always yield a positive solution. Thus, for this application, all monotonic schemes are naturally positive definite. While it is essential to have positive-definite schemes, that trait alone may be insufficient for air quality modeling and monotonic schemes may be more desirable. Advection schemes fall under one of two categories: low-order schemes and high-order schemes. It is well known that the low-order schemes display considerable numerical diffusion which can easily outweigh physical horizontal diffusion. O n the other hand, higher-order schemes generate spurious oscillations. Most schemes try to compromise between the dissipation error (or numerical diffusion) and oscillations related to dispersion (or phase shift) errors and the Gibbs phenomenon. The only way to minimize one of these errors without significant increase of the other is to introduce a nonlinear mechanism (Godunov, 1959). Typically, this mechanism is in the form of nonlinear flux corrections or nonlinear filtering. In advection schemes such adjustments are either applied implicitly or explicitly as a subsequent step to the linear solution. Advection schemes with different properties introduce different errors, all of which are sources of uncertainty in air quality model predictions. It is critical to identify which of the above-mentioned properties a scheme possesses before recommending its use. Since an advection scheme with all the desired properties is currently not available, the problem becomes identifying the scheme with the most desirable properties and eficiency. 9.2

ADVECTION SCHEMES

9.2.1 Classification of Schemes

There are many different ways of classifying advection schemes. A common way is to classify the schemes based on the method used in their formulation (Rood, 1987). Since a wide variety of methods are used, any classification may fall short of being complete. The following is a fairly comprehensive list: (1) finite difference schemes, (2) finite volume schemes, (3) flux corrected schemes, (4) Lagrangian schemes, (5) finite element schemes, and (6) spectral schemes. Current trends in advection scheme development show a merging of the methods to take advantage of each approach’s most desirable properties. For example, the characteristicGalerkin method (Childs and Morton, 1990) combines the best of the finite

346

Odman and Russell

element and Lagrangian methods. Flux corrections are being used in the framework of finite element and spectral schemes (Lohner et al., 1987). Also, the classical finite difference schemes are being abandoned in favor of modern finite volume schemes. Finite Difference Schemes

Finite difference schemes are general approximation techniques to differential equations. The derivatives are approximated by differences such that differential equations are transformed into algebraic equations and continuous variables are represented at discrete points. The fifth-order compact upwind differencing scheme (Tolst ykh, 1994) represents the state-of-the-art in finite difference schemes. It is mass conservative and it minimizes numerical diffusion and phase-shift errors. Also, the compact differencing operator makes the scheme more efficient compared to one that uses a full fifth-order differencing operator. The third-order Adams time integration described in Tolstykh (1994) requires information at several time steps. Multistep time integration methods are not feasible in air quality models because operators other than advection (chemistry, vertical diffusion) are applied between successive applications of the advection operator. A Runge-Kutta algorithm may also be used for time integration, but the resulting scheme does not provide a positive-definite solution and is computationally expensive. Therefore, although this scheme seems very accurate for moisture transport in general circulation models, it is not optimal for air quality modeling. Finite Volume Schemes

Here, we focus on a particular family of schemes classified as “volume schemes” in Rood (1987). These schemes assume a piecewise continuous concentration distribution. While the concentration is continuous within each volume element or grid cell, there may be discontinuities at cell interfaces. The fluxes through the faces of the volume element are computed by integrating the subgrid distribution of concentration between the cell face and the departure point of the last particle that would leave the cell at the end of the time step. Tremback et al. ( 1987) used higher-order area-preserving polynomials (e.g., a quartic with coefficients computed using area-preserving constraints) to represent the subgrid distribution of the advected scalar. Bott (1989a) normalized Tremback’s advective fluxes in an attempt to reduce the phasespeed errors. Negative values of the transported quantity are suppressed by nonlinearly limiting the normalized fluxes. Recently, a monotonic version of the scheme was developed (Bott, 1992) and the time-splitting errors associated with the use of one-dimensional operators in multidimensional appli-

Air Quality Modeling

347

cations were reduced (Bott, 1993). A version of this algorithm with second-order polynomials for nonuniform grid spacing in the vertical also exists (Strand and Hov, 1993). These schemes are being increasingly used in atmospheric and air quality applications (Easter, 1993; Chlond, 1994). In the piecewise parabolic method (PPM), the subgrid distribution of the advected quantity is represented by a parabola in each grid interval (Colella and Woodward, 1984). The approach is different from global spline methods (Emde, 1992) that try to fit a continuous curve to the advected quantity. A global fit is usually not well suited for representing localized, sharp transitions such as pollutant puffs. PPM not only provides a local fit of the data, but is monotonic and uses a special steepening procedure in the vicinity of sharp gradients. This ensures that positive quantities will remain positive and that sharp gradients will be handled correctly without the generation of spurious oscillations (Carpenter et al., 1990). This method has gained wide acceptance in the field of computational fluid dynamics (Bell et al., 1994) and has been also evaluated for atmospheric modeling (Carpenter et al., 1990). In contrast to PPM, Emde (1992) developed a global method that uses a continuous curvature cubic spline method based on Purnell’s scheme (Purnell, 1976). Purnell’s scheme is a conservative, positive-definite algorithm that is second-order accurate in time and third-order accurate in space and has good amplitude and phase properties (Pielke, 1984). The positive-definite property is achieved by breaking the spline locally in grid cells where the concentrations become negative and fitting other curves (e.g., straight lines) in such a way that the mass is conserved. Yamartino (1993) used piecewise cubic interpolands as a starting point for his higher-order scheme. In this scheme, the coefficients of a cellcentered cubic polynomial are constrained from the point of view of maintaining high-accuracy and low-diffusion characteristics while avoiding undesirable by-products (e.g., negative concentrations) associated with higher-order schemes but absent in low-order schemes. In addition, a filter is used for filling in undesired short-wavelength minima. One advantage of Yamartino’s scheme is that it was designed to follow short-wavelength features (e.g., plumes from single grid point sources). This scheme has been specifically designed for air quality modeling. The use of finite-volume schemes for atmospheric and air quality applications is very well developed, and these schemes can provide good accuracy fairly efficiently. Bott’s scheme, Yamartino’s scheme, and PPM are all highly recommended and will be further described shortly. While Bott’s scheme is very simple and efficient, Yamartino’s scheme is highly accurate in following short-wavelength features, which is a very desirable property in air quality modeling. PPM is the only truly monotonic scheme among the

340

Odman and Russell

three, which is another desirable property. Further testing of these advection schemes is necessary to recommend any one of them over the others. Flux-Corrected Transport Schemes

Flux-corrected transport (FCT) is a technique developed by Boris and Book (1973, 1976) and Book et al. (1975). The net transport flux is calculated as a weighted average of a flux computed by a low-order scheme and a flux computed by a higher-order scheme. The weighting is done in a manner that ensures that the higher-order flux is used to the greatest extent possible without introducing ripples (overshots and undershots). Zalesak (1979) generalized FCT to multidimensions. The performance of the more recent finite-volume schemes in the previous subsection are usually better than the FCT schemes. Semi-Lagrangian Schemes

In a semi-Lagrangian method, one estimates the backward trajectory of a particle (or air parcel) that arrives at a certain grid point. Since the origin of a particle does not always coincide with a grid point, an interpolation scheme is necessary to estimate the original concentration. Once estimated, this concentration is assigned to the grid point of arrival. One advantage of the scheme is that it is not subject to the Courant stability condition, so large time steps can be used. However, the method is not mass conservative, and spurious oscillations may be generated (depending on the choice of interpolation scheme). Assuring mass conservation and obtaining positive definite or monotonic results requires special treatments. In the method described by Williamson and Rasch (1989) and Rasch and Williamson (1990), shape-preserving interpolands are used to maintain the monotonicity of the discrete data. Interpolation is unnecessary in the method described by Olim (1994), but it results in a scheme that is neither monotonic nor positive definite. Thus, of the semi-Lagrangian methods, Rasch and Williamson’s method may be more appropriate for air quality modeling. In addition, this method has found acceptance in the community climate model (Williamson and Olson, 1994) and has been extensively tested and evaluated in that context. Finite Element Schemes

The traditional Galerkin finite element approach uses a method of weighted residuals in which the solution is expanded in piecewise basis functions (e.g., chapeau functions for a one-dimensional problem). The residual is assumed to be orthogonal to the weighting functions. Pepper et al. (1979) indicated

Air Quality Modeling

349

that this method combined with Crank-Nicolson time integration scheme preserves the peak value of a pulse quite well, but creates ripples behind the pulse. Raymond and Garder (1976) added a dissipative method that suppresses these ripples although the peak is also suppressed. This scheme was found to perform well by Chock and Dunker (1983): it has no restrictions on the Courant number, gives relatively accurate results, and is fairly efficient. Chock (1985) analyzed several different variations of this scheme and found that chapeau functions with the Forester filter (Forester, 1977) appears to be the best variation. The use of special weighting functions that are different than the basis functions results in “upwinding” effects similar to those in finite difference methods. This technique is known as the Petrov-Galerkin method. Upwinding introduces some artificial diffusion that reduces spurious oscillations. Both Hughes and Brooks (1979) and Kelly et al. (1980) found that artificial diffusion, only in the direction of flow, was desirable and could be implemented, thereby eliminating the problem of crosswind diffusion. Hughes and Brooks (1979) and Brooks and Hughes (1982) proposed a formulation that modifies the usual weighting functions by a perturbation dependent upon the velocity field and the derivative of the basis functions. This scheme is called the streamline upwind/Petrov-Galerkin (SU/PG) method. Tezduyar et al. (1987) developed SU/PG for time-dependent advection-diffusion equations. Odman and Russell (1991a, 1991b) implemented SU/PG in an air quality model. Donea (1984) describes an alternative algorithm for advective transport problems with improved stability properties. The time derivative is expanded in Taylor series; thus this method is called Taylor-Galerkin. The second- and third-order time derivatives are evaluated from the original equation in a manner similar to that employed in Hirt’s heuristic stability analysis (Hirt, 1968). This process yields a generalized governing equation that is discretized in time only; the spatial variables are left continuous. Such an equation is successively discretized in space using the conventional Galerkin finite element method. One disadvantage of the Taylor-Galerkin method is that spurious oscillations are not completely suppressed. Parrot and Christie (1986) and then Lohner et al. (1987) linked this method with the concept of flux-corrected transport (Zalesak, 1979) and obtained a monotonic scheme. In this scheme, a low-order scheme (e.g., lumped mass Taylor-Galerkin scheme with added artificial diffusion) guaranteed to give monotonic results is combined with a higher-order scheme (e.g., consistentmass Taylor-Galerkin scheme). Chock (1991) evaluated the onedimensional Taylor-Galerkin method followed by a nonlinear filter. While Taylor-Galerkin is probably the best choice among finite element advection schemes, significant benefits are observed only when the scheme is used in

Odman and Russell

350

two dimensions (possibly with flux correction). Two-dimensional codes have appeared recently (Lohner et al., 1987; Childs and Morton, 1990), but they are not readily accessible. Therefore, the Taylor-Galerkin method is not included in this study. Pseudospectral Schemes

In a pseudospectral scheme the Taylor series expansion is used to replace the time derivative with space derivatives. Then fast Fourier transforms are used to approximate the space derivatives (Gazdag, 1973; Wengle and Seinfeld, 1978; Chock, 1991). This method is known as the accurate space derivative (ASD) scheme. This scheme is highly accurate; however, it needs to be coupled with a nonlinear filter (such as the Forester filter) to suppress the spurious oscillations that may exist in the solution. A disadvantage of the scheme is the requirement of periodic boundary conditions inherent to all spectral schemes. However, there are several solutions for this problem (Gazdag, 1973; Roache, 1978; Wengle and Seinfeld, 1978; Chock, 1991). Since the boundary conditions in air quality models are not periodic, special treatments are required at the boundary. The pseudospectral scheme was found to be the most accurate of several methods tested by Chock (1985), but it required a long execution time. Chock (1991) compared several chapeau function methods, and ASD and found that the chapeau function with Taylor-Galerkin and a Forester filter was an excellent choice if one was concerned with execution time. However, the computationally intensive ASD method was significantly more accurate. Thus, the choice of advection scheme depends on the accuracy desired and the computational tools available. Recently, the parameters of the Forester filter were optimized for the ASD scheme by Dabdub and Seinfeld (1 994). 9.2.2 Brief Description of Schemes

Generally, air quality models solve the advection-diffusion-reaction equation using the time-splitting approach (Yanenko, 1971). Each process is split into one-dimensional operators and the solution from one process is used as the initial condition for the next process. The advection of a tracer species is represented by the following conservation equation :

- + v - (uc) = 0 8C

(9.15)

where c is the species concentration (in density units) and vector.

is the velocity

Air Quality Modeling

351

Time splitting leads to the following one-dimensional equation in Cartesian coordinates : ac a(uc) -+-at ax - 0

(9.16)

To simplify the discussion, we consider a uniform (i.e., constant A x ) and staggered grid (cj represents the grid cell average of the concentration, while u j + 1/2 is the advection velocity defined at grid cell interfaces). In thc finite-volume schemes discussed below (the piecewise parabolic method, the Bott scheme, and the Yamartino scheme), we use the explicit flux formula to advance the solution a single time step, At, from level n to level n + 1:

where F;+l/2 and Fj"-1/2are the advective fluxes of c through the right and left boundaries of grid cell j , respectively. Further, let us define a nondimensional coordinate x as = (x - xj)/Ax so that, in grid cell j , -1/2 _< ( I1/2. Now, suppose that the concentration has a certain distribution c,(t) in each grid cell. Depending on the direction of the velocity, the flux F j + 1/2 can be expressed as CAX

(9.18)

where c j + 1/2 = I u j + 1/2 I At/Ax is the Courant number at the right boundary of grid cell j . The departure point of the last particle that crosses the cell interface appears in the limits of the flux integrals as tdp= 1/2 T c j + 1 / 2 . In calculating this expression it was assumed that the velocity is uniform over the grid cell. Assuming linear variation of the winds between the two cell faces may lead to a more refined estimation of the mass fluxes through the cell faces. The computation of the departure point for linearly varying velocity can be found, for example, in Seibert and Morariu (1992). A practical correction for linearly varying fields is to build the divergence into the Courant number as & = &

(

I---

E:)

(9.19)

Earlier, we mentioned that the conditions of higher-order accuracy and freedom from spurious oscillations cannot be achieved simultaneously. The

Odman and Russell

352

only way to satisfy one of these conditions without significant violation of the other is to introduce a nonlinear mechanism. Typically, this mechanism is provided by nonlinear flux corrections or nonlinear filtering. In the advection schemes below, such adjustments are either applied implicitly through the solution or explicitly as a subsequent step to the linear solution. It is important to realize that the original problem described by Eq. (9.12) has a linear nature. Introduction of nonlinear mechanisms in order to improve certain aspects of the numerical solution modify this linear nature. The schemes that will be described here were recently tried in atmospheric modeling. They are (1) the Smolarkiewicz scheme, (2) the piecewise parabolic method, (3) the Bott scheme, (4) the Yamartino scheme, ( 5 ) fluxcorrected transport, (6) the semi-Lagrangian method, (7) the chapeau function scheme and Forester filter, and (8) the accurate space derivative (ASD) scheme.

Smolarkiewicz Scheme

The Smolarkiewicz scheme (Smolarkiewicz, 1983) is based on the first-order accurate upstream or “donor cell” method. In the donor cell method, the distribution of the advected quantity is assumed to be constant over the grid cell (i.e., c j ( c k - 1 , Fk-3f2 > Fk-1/2, and a ( c k - 2 - c k - l ) > ( F k - 3 / 2 - F k - l j Z ) l * Instead of increasing the underestimated flux, the limiter reduces the advective flux downwind, Fk+1 1 2 , in order to avoid an undershot in cell k. This eventually reduces the net flux out of the domain, resulting in an accumulation of mass in the domain. Flux -COrrec ted Transport Sch erne

The FCT procedure of Zalesak (1979) can be summarized as follows: Compute Ff+ 1/2, advective fluxes through cell faces given by a loworder scheme guaranteed to give monotonic solutions. We used the donor cell scheme as the low-order scheme. 2. Compute FY+1,2,advective fluxes through cell faces given by a higher-order scheme. We used fourth-order area-preserving polynomials (i.e., Bott's scheme except the flux limiting) as the higherorder scheme. 3. The antidiffusive flux, A j + 1 , 2 , is defined as the difference of higherand low-order fluxes: 1.

Compute the low-order solution as (9.30)

5. Limit the antidiffusive flux, Aj+ l j 2 , with a limiting factor, C j +1 1 2 , in a manner such that the solution c J + l as computed below is free of extrema not found in cy or cf: A:+

112

= Cj+lizAj+1 / 2 7

0I Cj+112 I 1

(9.31)

Air Quality Modeling

357

6. Compute the solution using limited antidiffusive fluxes: c;+l

= c4 - At ( Acj + l i Z - A?-I,,)

(9.32)

Although its details are omitted here, the limiting of the antidiffusive flux is the most important step. Detailed discussion of this can be found in Zalesak (1979). The accuracy of the flux-corrected schemes highly depends on the accuracy of the higher-order scheme employed. The FCT concept has also been implemented in finite element methods. Lohner et al. (1987) linked this concept with the Taylor-Galerkin method of Donea (1984). In this scheme, element contributions from a low-order scheme (e.g., lumped-mass Taylor-Galerkin scheme with added artificial diffusion) guaranteed to give monotonic results are computed first. Then the element contributions from a higher-order scheme (e.g., consistent-mass Taylor-Galerkin scheme) are computed. Note that because the finite element method is being used, the term “flux” in Zalesak’s original concept is replaced by “element contribution” here. In the next step, the antidiffusive element contributions are computed by taking the difference of higher-order and low-order element contributions. Finally, a low-order solution is computed, the antidiffusive element contribution is limited for monotonicity, and the low-order solution is updated by applying the limited antidiffusive element contribution. To our knowledge, this scheme has not been used in any atmospheric models. Yarnartino’s Scheme

Yamartino’s scheme (Yamartino, 1993) is another finite-volume scheme where the interpolating polynomial is a cubic spline: (9.33) a, = cj a, = dj A x

(9.34) The spline derivatives, dj, are obtained from the tridiagonal system: adj-

+ (1 - 2a)dj + adj+

Cj+ 1

- Cj-

2 Ax

(9.35)

Odman and Russell

358

with a = 0.22826. Note that a value of a = 0 would correspond to the explicit expressions of d j . The positivity of cjc) is ensured by various mechanisms. First, when c j is a local minimum, the donor cell scheme is used instead of the cubic spline. Second, the spline is spectrally limited by the relation (9.36)

where k = 1,2, 3. Third, a mass conservative flux renormalization is applied where the fluxes are normalized with the ratio for the upwind cell of the cell concentration (i.e., concentration at t = 0) divided by the average concentration. Finally, a mildly diffusive filter is applied in an attempt to block the depletion of donor cells. Yamartino’s scheme is not monotonic and can generate new maxima. Chapeau Function Scheme

The chapeau function method is a classical weighted-residual finite element method. The solution is expanded in piecewise basis functions, N k ( X ) , that look like hats (“chapeau” is French for hat) in one-dimensional space: P

(9.37) k=l

The residual is assumed to be orthogonal to the weighting functions, which may be the basis functions themselves, as is usually the case in the BubnovGalerkin methods: (9.38)

After a coordinate transformation 5 = (x - x j ) / A xwhere - 1/2 5 c 5 1/2, the basis functions of a typical finite element with two nodes can be written as (9.39)

and the weighted residual equation becomes (9.40)

Air Quality Modeling

359

M i j -+ K i j c j = 0 at

(9.41)

where

for each finite element. If the Crank-Nicolson time-integration is used then Eq. (9.41) becomes (9.43)

This equation results in the following tridiagonal system:

The numerical diffusion introduced by this scheme is quite small, but spurious oscillations are observed near steep gradients (Pepper et al., 1979). McRae et al. (1982) applied a smoothing filter to suppress these oscillations. This mass-conservative filter locally and selectively damps the shortwavelength noise by introducing artificial diffusion (Forester, 1977). Due to the local property of the filter, resolvable features of the linear solution such as peak pollutant concentrations are maintained. The filter can be applied successively until positive-definite results are obtained. Recently, Odman and Russell(l993) enhanced the performance of the Forester filter and generalized it for use with multidimensional finite element schemes. Accurate Space Derivative Scheme

The accurate space derivative (ASD) scheme (Chock, 1991; Dabdub and Seinfeld, 1994) is classified as a spectral scheme and is highly accurate. However, it needs to be coupled with a nonlinear filter, such as the Forester filter, to suppress the spurious oscillations that may exist in the solution. Also, special treatments are necessary to satisfy the periodic boundary condition requirement since the boundary conditions in air quality models are

360

Odman and Russell

not periodic. Computationally, the scheme is generally more expensive than others, but its high accuracy makes it particularly attractive. First, the concentration is expanded as a truncated Taylor series in time. The order of accuracy of this explicit scheme can be increased by including more terms in the series. Here, we include terms up to the thirdorder time derivative: (9.45)

Then the time derivatives of the concentration are replaced by the space derivatives using the governing equation (i.e., advection equation). Note that the first time derivative already appears in this equation. Higher-order derivatives can be obtained by differentiating this equation. Assuming that the velocity is not a function of time, the spatial equivalents of time derivatives can be written in conservation law form, as (7c _ -

c72C

d(UC)

-?X

3t2 - 8x

d(uc)

x] (9.46)

However, to keep the derivatives as simplified as possible for the fast Fourier transform (FFT), the derivatives are computed in chain rule form: au - ac ac -- -c _ U

8.x

Next, the space derivatives of both c and U are obtained through the use of an FFT. To comply with the periodic boundary conditions requirement of the FFT, we add six more cells to the domain beyond the real boundaries. Then we create data for these extra cells utilizing a polynomial fit and forcing periodicity of the data and the derivatives. The polynomial is either parabolic or cubic, depending on the signs of the derivatives at the boundaries. If the derivatives at the boundaries are of opposite sign, then

Air Quality Modeling

361

we use a parabolic fit; if the derivatives at the boundaries are of the same sign, we employ a cubic fit. Once the periodic boundary conditions are enforced, we transform the data to the Fourier space with forward FFT, take the derivatives in that space, and then transform the derivatives back to the original space using backward FFT. The computation of the derivative in the Fourier space assures high-order accuracy throughout the entire domain, including grid cells near the boundaries. We then substitute the derivatives into Eq. (9.45) and update the concentration field. Due to its dispersive quality, the F F T approach tends to produce highfrequency noise. This noise is filtered and diffused using the nonlinear Forester filter :

where is a weight function determining the amount of filtering applied to cellj, and K, is the coeficient of artificial diffusion. The filter is designed to be used in conjunction with second- and higher-order methods. Computational noise is minimized without incurring amplitude penalty. This is achieved by being able to select the desired frequencies of noise and adding diffusion locally at selected grid cells at which x j is set equal to one and held at zero elsewhere. This way, the high-frequency noise generated by the F F T can be diffused away. Semi-Lagrangian Transport Scheme

In the semi-Lagrangian transport (SLT) scheme, one estimates the backward trajectory of a particle (or air parcel) that arrives at a certain grid point. Since the origin of a particle does not always coincide with a grid point, an interpolation scheme is necessary to estimate the original concentration. Once estimated, this concentration is assigned to the grid point of arrival. One advantage of the scheme is that it is not subject to the Courant stability condition, so large time steps can be used. However, the method is not mass conservative, and spurious oscillations may be generated (depending on the choice of interpolation scheme). Assuring mass conservation and obtaining positive definite or monotonic results requires special treatments. There are several semi-Lagrangian schemes listed in the literature (Rasch and Williamson, 1990; Smolarkiewicz and Grell, 1992). SemiLagrangian methods are being used particularly in global circulation and climate models (Williamson and Rasch, 1989). The SLT scheme is based on the advective form in Eq. (9.13): 84

84 ax

--+U-+v-=o

84 ay

(9.49)

Odman and Russell

362

where q is the mass mixing ratio related to the concentration c and density of air p as q = c/p. The Lagrangian solution to this equation determines the departure point ( x D ,y D )of a particle at ( x A ,y A )as (9.50) The SLT scheme of Rasch and Williamson (1990) is described here. This scheme first determines the midpoint of the trajectory iteratively as

.&+

= ?CA - -ju ( x L , y”,)

(9.51)

Four iterations are used for the very first time step, which starts with the arrival points as a first guess (for the midpoints) and one iteration thereafter where the midpoints from the previous time step are used as a first guess. The velocities at the midpoints are calculated using Lagrange cubic interpolation. The departure points are calculated as (9.52) Finally, the q field at the departure points are found by Hermite cubic interpolations and used to update the values at arrival points: q(t) = (2t3 - 3 t 2 + 1)qj + (-2t3 + 3t2)qj+1 + (t3- 2t2 + t)dj + (t3- t2Vj+1

(9.53)

where (9.54) The arrival points are grid points. The reason for using the Hermite cubic interpolation is that the derivative terms are explicit and they may be limited to obtain monotonic results. A sufficient (but not necessary) condition for monotonicity is (9.55)

Air Quality Modeling

363

where Aj is the discrete slope defined as Aj = (qj+l - qj)/Ax. Also, instead of using two-dimensional bicubic interpolants, the tensor products are employed where a series of interpolations in one dimension are followed by one interpolation in the other dimension. According to which direction is interpolated first, one may get a different answer, but the differences are minor according to Williamson and Rasch (1989). 9.3

EVALUATION OF ADVECTION SCHEMES

9.3.1 Description of Tests and Performance Measures Used

Typically, the performances of advection schemes are measured and compared with each other using test cases with idealized flow fields. These ideal flow tests have analytic solutions and are very useful for determining certain properties of the schemes. Here, we perform a preliminary evaluation to identify schemes with more desirable properties. As discussed earlier, the properties desirable in an advection scheme for use in photochemical modeling are mass conservation, small numerical diffusion and phase-speed errors, and freedom from Gibbs oscillations at least to the extent where they do not lead to negative concentrations and monotonicity. In the figures and tables in this section, the schemes are abbreviated as follows : ASD BOT BOT-M BOT2D FCT HAT PPM SLT SMO YAM

accurate space derivative scheme Bott’s scheme Bott’s scheme with the monotonic limiter two-dimensional version of Bott’s scheme flux-corrected transport scheme chapeau function scheme piecewise parabolic method semi-Lagrangian transport scheme Smolarkiewicz’s scheme Yamartino’s scheme

The following tests are used for preliminary evaluation of the schemes: 1. One-dimensional tests: Tests where pulses of different shapes are advected with uniform velocity. 2. Rotating cone test: A cone-shaped puff is introduced in a rotational flow field and followed for a certain number of revolutions. 3. Skew advection of a point-source plume: The plume from a continuously emitting point source is advected diagonal to the grid. 4. Shear Jlow tests: The effects of wind shear are simulated with a stagnation or a vortex flow field.

Odman and Russell

364

5. Rotatinglreacting cones: Cone-shaped puffs of NO, and ROGs are rotated while they form ozone through a simplified nonlinear photochemical mechanism.

We based the evaluation and comparison of the schemes on the following performance measures : 1. Peak ratio: This ratio is a measure of the peak retention capability of an advection scheme. A value of 1.00 means that the scheme has perfect peak retention. The numerical diffusion introduced by some schemes may result in a peak clipping effect, resulting in a peak ratio that is smaller than unity. The ratio can also be larger than unity for some schemes that overshoot peak concentrations. The peak ratio is calculated here as

max(c,) max(cf) ’

i = l , 2 , 3 ,..., n

(9.56)

where cf is the exact solution for grid cell i, ci is the predicted value, and n is the number of cells. 2. Background-to-peak ratio: In all tests, the background concentration is initialized to a nonzero value. If a scheme is not monotonic, it will create ripples near steep gradients. The ripples lead to predicted concentration values smaller than the background. T o measure the magnitude of these ripples, the background-to-peak ratio is computed. The amplitude of the ripples is compared with the amplitude of the signal that should be resolved (e.g., peak concentration). If the ratio is negative, this means that the scheme is not positive definite. The perfect value for the background-to-peak ratio may vary from one test to another; we will state it for each case. The ratio is calculated as min(c,) m a x(cf )

(9.57)

3. Mass ratio: This ratio is a measure of the mass conservation capability of the schemes. A value of 1.00 means that the scheme conserves mass. Nonconservative schemes may lose (mass ratio smaller than unity) or gain (mass ratio greater than unity) mass. Here, the mass ratio is calculated as (9.58) 4.

Distribution rutio: This ratio is a measure of how well the mass is distributed in comparison to the exact solution. A value of 1.00

Air Quality Modeling

365 Gaussian Wave PMCWppm) Layer 1 Step 50 Smolahiemcz Scheme

Exact Solullon 95

71.25

47 5

47.5

23.75

i v

100

Sleo

71.25

71 25

47.5

23.75

0 50

100

Chapeau Function Method

Semi-LagranyanTtanspon

Peak a1 (75)= 57.0356

Peak at (75)= 94.2442

50 St@P

slap

Peak a1 (75)= 46.8216

100

loo

8l.p

Peak a (75) = 74.1205

Figure 9.3 Solutions to the Gaussian wave test: exact solution (top left), SMO (top right), SLT (bottom left), and HAT (bottom right).

means that the scheme distributes the mass perfectly. The distribution ratio is calculated as (9.59)

Odman and Russell

366 G.uuunwaw PMCH(ppm) by.r 1 step 50

71.25

47.5

23.75

v 25

50 stao

loo

Peak U (75) = 942442

71.25

47.5

23.75

100

0 0

1OO

rsrp

Peak U (75) = 65 2109

Figure 9.4 Solutions to the Gaussian wave test: exact solution (top left), BOT (top right), BOT-M (bottom left), and PPM (bottom right).

Average absolute error: This error is calculated as (9.60)

Air Quality Modeling

367

and is an indicator of overall scheme performance. 6. Root mean square ( R M S )error: This error is calculated as (9.61)

9.3.2 Evaluation Cases Advection of One-Dimensional Pulses

In one-dimensional tests, pulses of various shapes are advected with constant velocity. Rectangular, triangular, and Gaussian-shaped pulses of inert species are very common (Muller, 1992). Usually, the one-dimensional tests cover a wide range of Courant numbers (a measure of how far the pulse will move in one time step compared to the grid size) for a full account of the stability and accuracy of the schemes. Traditional tests with longwavelength pulses tend to show an advection scheme at its best. Adequate short-wavelength performance is also extremely important in photochemical models but is usually more difficult to obtain. Here, a Gaussian wave of wavelength equal to 8 Ax is advected from cell 25 to cell 75 of a 100-cell uniform grid. Figures 9.3-9.5 show the shape of the Gaussian pulse after it has been advected a distance of 50Ax at a Courant number of 0.25 (i.e., after 200 time steps) as predicted by each scheme. Table 9.2 summarizes the values of various performance measures at the end of the test. The A S D and YAM schemes preserved the peak height very well. However, the distribution ratio has a lower value for YAM, indicating distortions of the pulse’s shape. For the same reason, the average and RMS errors are larger than those of ASD. On all accounts, these two schemes perform much better than the

Table 9.2

Gaussian Wave Test Scheme

Performance measure

ASD

BOT

BOT-M

FCT

HAT

PPM

SLT

SMO

YAM

Peak ratio Background Mass ratio Distribution Average error RMS error

0.99 0.05 1.00 0.99 0.08 0.01

0.87 0.01 1.00 0.93 0.87 0.18

0.74 0.05 1.02 0.83 1.38 0.27

0.74 0.05 1.00 0.81 0.98 0.14

0.79 0.05 1.00 0.77 1.01 0.20

0.69 0.05 1.00 0.79 1.16 0.17

0.50 0.05 0.95 0.52 2.13 0.40

0.61 0.02 1.00 0.66 2.21 0.50

0.98 0.05 1.00 0.92 0.51 0.12

Odman and Russell

368 Gauuun Wavr PMCH(ppm) Layrr 1 Slap50

71 25

71.25

47.5

23.75

100

slop P&

a1 (75)

P W .f (76) = 93.1378

94.2442

VunarWdr Schema 95

71.25

47.5

23.75

0 0

50 slop

loo

u.0

loo

P u k at (76) = 924035

Figure 9.5 Solutions to the Gaussian wave test: exact solution (top left). ASD (top right), YAM (bottom left), and FCT (bottom right).

other schemes in this test. Bott’s scheme (BOT) ranks third overall, but large ripples are observed at leading and trailing edges of the pulse as indicated by the values below the background (as much as 4% of the peak height). When the monotonic limiter is used (BOT-M), the ripples are

Air Quality Modeling

369

Figure 9.6 Solutions to the rotating cone test: exact solution (top left), SMO (top right), SLT (bottom left), and HAT (bottom right).

eliminated but the peak retention performance deteriorates. Also, a 2% increase in mass is observed. The FCT scheme has the same peak ratio as BOT-M while it conserves mass and leads to smaller errors than BOT-M. The chapeau function scheme (HAT) produces a higher peak than PPM and BOT-M, but the shape of the puff is not maintained as well; this is also indicated by a smaller distribution ratio. Smolarkiewicz’s scheme (SMO) and SLT display poorer performance than other schemes. SMO produces ripples upwind from the pulse and leads to average and RMS errors larger than the SLT scheme’s. However, 5% of the mass is lost with SLT and the peak retention is much worse than with SMO.

370

Odman and Russell

Figure 9.7 Solutions to the rotating cone test: exact solution (top left), BOT (top right), BOT-M (bottom left), and PPM (bottom right).

Rotating Cone Test

The rotating cone test is a routine way of measuring the long-wavelength advection fidelity of the schemes (Crowley, 1968; Chock, 1991; Odman and Russell, 1991b; Odman et al., 1995, 1996). In this test, a cone-shaped puff is introduced into a rotational flow field. The exact solution to this problem is a rigid-body rotation of the puff without any change to its original conical shape. Various errors are revealed in this test. For example, the numerical diffusion (or dissipation) error manifests itself in the drop of the peak height

37 1

Air Quality Modeling

during rotation. Also, by observing the location of the peak, one can determine the leading or lagging phase-speed errors. A 32 x 32 grid is used for this test (i.e., - 16Ax I x 5 + 16Ax; - 16Ay 5 y 5 + 16Ay; Ax = Ay). A cone-shaped puff with peak concentration equal to 100 ppm and a base radius of 4Ax is initialized such that its peak is located at [ + 8Ax, 01. Note that the peak is not initially at a gridcell center but at a cell corner (i.e., there are four cells around the peak with the same average concentration). The background concentration is set equal to 5 ppm. To obtain a counterclockwise rotation around an axis passing through the center of the domain, the wind field is defined as follows : U =

-my,

u=ox

(9.62)

The angular velocity, o,is adjusted so that the Courant number is approximately 0.28 at the peak of the puff. Figures 9.6-9.8 show the solution after two full rotations as predicted by each scheme. The exact solution, which is nothing but the initial cone, is also shown. ASD and YAM maintain the peak height and the overall shape of the cone better than the others. BOT performs third best in this test, but it yields values below the background (as seen in the ring-shaped valley at the base of the cone). BOT-M, HAT, and PPM predict similar peak heights (65%, 7070,and 60%, respectively) but the shape distortions look very different in each case. PPM has the worst peak-clipping effect but the resulting shape has the smallest base span among the three schemes. SLT and SMO are clearly the most diffusive schemes; SMO also introduces a ripple upwind from the cone. Table 9.3 summarizes the performance measures at the end of two rotations. Since there is no shear in the flow field, BOT-M and the two-dimensional version of Bott’s scheme (BOT2D) (Bott, 1993)

Table 9.3 Rotating Cone Test Scheme

Performance measure

ASD

BOT

BOT-M

FCT

HAT

PPM

SLT

SMO

YAM

Peak ratio Background Mass ratio Distribution Average error RMS error

0.99 0.06 1.00 0.96 0.18 0.05

0.87 0.03 1.00 0.93 0.46 0.16

0.65 0.06 1.02 0.83 0.76 0.30

0.69 0.06 1.00 0.80 0.48 0.14

0.71 0.06 1.00 0.70 0.79 0.22

0.61 0.06 1.00 0.78 0.54 0.18

0.32 0.06 0.94 0.45 1.31 0.31

0.49 0.02 1.00 0.64 1.60 0.51

0.99 0.06 1.00 0.91 0.33 0.13

372

Odman and Russet1

Figure 9.8 Solutions to the rotating cone test: exact solution (top left), ASD (top right), YAM (bottom left), and FCT (bottom right).

produce identical results. The mass conservation problem with these two schemes is revealed once again. SLT does not conserve mass either, with 6% of the mass lost. BOT preserves 87% of the peak height (third best after ASD and YAM), but it leads to ripples with an amplitude of 3% of the original peak height. FCT solution is free from ripples while the average and RMS errors are of the same order as BOT, but it has a peak ratio of only 69%. The performance of BOT-M, HAT, and PPM are comparable with HAT, predicting the highest peak but producing the lowest distribution ratio, and PPM gives the smallest average and RMS errors among the

Air Quality Modeling

373

Figure 9.9 Solutions to the skew advection of a point-source plume test: exact solution (top left), SMO (top right), SLT (bottom left), and HAT (bottom right).

three. Notice that the comparison results obtained from this test are very similar to those of the Gaussian wave test above. Skew Advection of a Point-Source Plume

Advection of a point-source plume is a test problem specifically designed to measure the short-wavelength performance of the schemes (Yamartino, 1993; Odman et al., 1995). In photochemical modeling, there is often a need to simulate the transport of plumes emitted into a single-grid cell. In this

374

Odman and Russell

Figure 9.10 Solutions to the skew advection of a point-source plume test: exact solution (top left), BOT (top right), BOT-M (bottom left), and PPM (bottom right).

test, a point source located in a single-grid cell (cell [5,5] of a 32 x 32 grid) emits at a constant rate and the winds advect the plume diagonally (i.e., skew advection). The emission rate is adjusted to make up for the advected mass as predicted from the analytical solution, so that the concentration in the source cell remains constant at 100 ppm. All other cells are initialized with a background concentration of 5 ppm. The wind field is defined such that the plume is advected along the diagonal of the domain with a Courant number of 0.17 (0.12 in each coordinate direction). Ideally, the concentration along the diagonal should be constant and equal to the con-

Air Quality Modeling

375

centration in the source cell (i.e., 100 ppm). This stringent test also reveals how much crosswind diffusion is introduced by the schemes. The plume as predicted by each scheme is shown in Figs. 9.9-9.11 at the instant when it reaches cell [28,28] according to the exact solution. A mass conservation problem is evident in the HAT solution. This was related to a smoothing treatment of the concentrations near the boundaries. After correcting this problem, however, we were still unable to get a stable solution from the HAT scheme even after trying various filtering parameters. Table 9.4 summarizes the results of this test for all the other schemes. It

Figure 9.11 Solutions to the skew advection of a point-source plume test: exact solution (top left), ASD (top right), YAM (bottom left), and FCT (bottom right).

376 Table 9.4

Odman and Russell Skew Advection of a Point-Source Plume Scheme

Performance measure

ASD

BOT

BOT-M

FCT

PPM

SLT

SMO

YAM

Peak Background Mass Distribution Average error RMS error Cell [15,15] Cell [25,25]

0.71 0.03 0.98 0.62 1.81 0.47 0.68 0.67

0.98 0.00 1.00 0.72 2.51 0.87 0.62 0.56

0.95 0.05 1.04 0.62 2.98 1.13 0.46 0.43

0.94 0.05 1.00 0.58 2.91 1.01 0.45 0.39

1.00 0.05 1.00 0.57 3.10 1.08 0.42 0.38

1.25 0.05 1.16 0.79 4.30 1.54 0.46 0.38

1.21 0.02 1.00

0.97 0.05 1.00 0.72 1.39 0.49 0.72 0.62

0.55 3.64 1.05 0.44 0.35

also gives the ratio of predicted solutions to exact solution at cells [lS,lS] and [25,25], which are approximately 14Ax and 28Ax downwind from the source. Note that, as in the rotating cone test, there is no shear in the flow field, so the results for BOT-M and BOT2D are the same and only BOT-M is shown. In this test, YAM performs better than A S D due to its special shortwavelength capturing features. A S D underpredicts the peak height near the source, but further downwind the peak retention becomes better than all other schemes. At cell [lS,lS], ASD’s peak ratio is second only to YAM’s, but it is the highest (67%) at cell [25,25]. Mass conservation problems are seen with A S D (2% loss), BOT-M (4% increase), and SLT (16% increase). The average and RMS errors are smallest for YAM and ASD. SLT has the highest distribution ratio, but this is due to the added mass, so judging its performance based on this measure alone would be misleading. Also, note that the errors for SLT are larger than for any other scheme and that the peak is overpredicted by 25%. SMO also overpredicts the peak near the source by 21% and leads to the lowest concentration values further downwind. The distribution ratio is low and the average error is high due to this incorrect distribution of the mass in the plume. BOT predicts a peak concentration higher than both ASD’s and YAM’s near the source, but the values are much smaller downwind. Also, a ripple with an amplitude equal to 5% of the exact plume height is produced upwind from the source. However, the concentrations in this ripple do not go below zero due to the positive-definite nature of BOT. S M O produces a ripple of smaller magnitude (the amplitude is only 3% of the exact peak concentration). BOT-M does not lead to ripples, but achieves this at the expense of added artificial diffusion as seen in the reduced values of concentrations downwind. The

Air Quality Modeling

377

Figure 9.12 Solutions to the stagnation flow test: exact solution (top left), SMO (top right), SLT (bottom left), and HAT (bottom right).

performance of FCT is very close to BOT-M in this test. P P M performed sixth best, but it is the only scheme with perfect prediction of the peak concentration at the source. Shear Flow Tests

The variability of wind speed in actual wind fields is another factor that may affect the performance of advection schemes. Staniforth et al. (1987) and Odman and Russell(l993) designed robust test cases and demonstrated that some advection schemes may perform poorly under more realistic

378

Odrnan and Russell

Figure 9.13 Solutions to the stagnation flow test: exact solution (top left), BOT (top right), BOTZD (bottom left), and PPM (bottom right).

meteorological conditions. Here, the effects of wind shear are simulated as follows: A square-shaped puff is released in a stagnation flow field, and the field is reversed (or rotated by 90') each time the puff reaches an aspect ratio of 2 : 1. In other words, the velocity field is u=ox,

U =

-oy

(9.63)

during the first and last quarters of a cycle, and U = -ox,

u=oy

(9.64)

Air Quality Modeling

379

during the second and third quarters such that the puff is first stretched in the x-direction, then in the y-direction and finally in the x-direction until the original square shape is exactly restored at the end of each cycle. The solution at intermediate stages can be found in Seibert and Morariu (1991) or Fogelson (1992). The solutions to the stagnation flow problem as predicted by each scheme after three cycles are shown in Figs. 9.12-9.14 and the performance measures are listed in Table 9.5. In this test, the numerical diffusion of most schemes leads to mass loss through the boundaries when the puff is fully

Figure 9.14 Solutions to the stagnation flow test: exact solution (top left), ASD (top right), YAM (bottom left), and FCT (bottom right).

380

Table 9.5

Odman and Russell

Stagnation Flow Test Scheme

Perform a nce measure

ASD

BOT

BOT2D

FCT

HAT

PPM

SLT

SMO

YAM

Peak Background Mass Distribution Average error RMS error

1.02 0.05 0.93 0.58 3.09 0.77

0.85 0.03 1.00 0.61 4.34 1.46

1.40 0.05 1.00 0.74 3.77 1.11

0.71 0.05 0.97 0.54 4.31 1.41

0.88 0.05 0.96 0.54 3.98 1.09

0.73 0.05 0.97 0.53 4.30 1.34

0.91 0.05 1.12 0.71 4.98 1.80

0.85 0.04 0.94 0.46 4.96 1.23

0.85 0.05 0.96 0.58 3.68 1.24

stretched (aspect ratio of 2 : 1). Due to this mass loss, some performance measures (e.g., mass ratio, distribution ratio, errors) may be misleading and should be used carefully in the evaluation. ASD slightly overpredicts the peak (2%), some ripples are observed in the background (Fig. 9.14), and 7% of the mass is lost. However, the average and RMS errors are smaller than with other schemes and the shape of the puff is in relatively good agreement with the exact solution. The performances of YAM and HAT are very close, but we rank HAT the second best performer in this test. Only 12% of the peak height and 4% of the mass are lost. YAM underpredicts the peak height by 15%, and loses 4% of the mass through the boundaries. Its average error is smaller than HAT’S, but its RMS error is larger. BOT is the fourth best performer, followed by PPM. BOT introduces ripples with amplitudes as much as 2% of the peak height, and P P M loses 27% of the peak height. The performance of FCT is almost identical to PPM’s. The solutions of BOT2D, SLT, and SMO do not compare well to the exact solution. BOT2D overpredicts the peak height by 40%, SLT increases the mass by 12%, and the SMO solution looks very diffusive. Rotation of Reactive Cones

The advection tests described so far investigate certain measures such as the degree of preservation of the peaks and the generation of ripples. There are some properties of transport schemes, however, that an advection test alone may not reveal. The nonlinear chemistry alters the shape of a pollutant puff constantly; it can change its slope, making it less or more steep, or even invert a peak. The ability of the scheme to adapt to such changes can be seen and evaluated only in tests with chemistry. Hov et al. (1989) demonstrated that the aliasing errors inherent to the highly accurate pseudospectral scheme may be amplified in the presence of nonlinear chemistry. Odman and Russell (1991b), in a similar test, showed that sacrificing mass

Air Quality Modeling

38 1

Figure 9.15 NO, predictions in the rotating/reacting cones test: exact solution (top left), SMO (top right), SLT (bottom left), and HAT (bottom right).

conservation may have important unfavorable consequences in modeling regional oxidant formation. For testing purposes, the chemical mechanism is kept simple; this way, sound physical arguments can be made more readily. However, the mechanism still yields the kind of numerical difficulties encountered in photochemical models. One such description of the photochemistry was used by Hov et al. (1989) and later by Odman and Russell (1991b) and Chock and Winkler (1994a, b). The test consists of advecting puffs of different species while they react according to the simple chemical mechanism. The solution

382

Odman and Russell

Figure 9.16 NO2 predictions in the rotating/reacting cones test: exact solution (top left), BOT (top right), BOT-M (bottom left), and PPM (bottom right).

to the advection part is known: the concentration field is simply rotated. O n the other hand, if no advection is taking place but the chemistry is activated, the concentrations will change and their values may be determined numerically by a very accurate method such as the Gear solver. If both advection and chemistry are activated, then the solution is simply a rotation of the fields one gets from performing the chemical integration. This solution (numerical chemistry and exact advection) is compared to the predictions from the advection schemes followed by chemistry. Four species (HC, HCHO, NO, and NO2) are initialized as cones with the same

Air Quality Modeling

383

geometry as in the rotating cone test above. The initial peak heights and background concentrations can be found in Odman and Russell(1991b). Since the shapes of different pollutant puffs are altered differently by the chemistry, the errors are usually not the same for all species. In this simple mechanism, the errors for species such as NO, NO,, HCHO, and 0, are usually more pronounced. This test is very reliable in studying how the errors produced by the advection schemes would propagate in the model. The results for NO, and 0, after 48 h of simulation are shown in Figs. 9.15-9.20, and some performance measured for HC, NO2, and 0, are summarized in Table 9.6. Note that all schemes with the exception of SMO and

Figure 9.17 NO, predictions in the rotating/reacting cones test: exact solution (top left), ASD (top right), YAM (bottom left), and FCT (bottom right).

384

Odman and Russell

Figure 9.18 0, predictions in the rotating/reacting cones test: exact solution (top left), SMO (top right), SLT (bottom left), and HAT (bottom right).

SLT predict the HC peak within a few percent. BOT overpredicts this peak by 2.5% while SMO and SLT underpredict it by 7% and 9%, respectively. Also, the RMS error for SMO and SLT are the largest. ASD, YAM, and FCT perform better than other schemes in following the HC puff. The NO, valley is best predicted by SMO (100%) but the shape of the solution in Fig. 9.15 and the large RMS error show that this may be fortuitous. The shape is severely distorted near the background due to the amplification of ripples in the presence of nonlinear chemistry. BOT produces similar ripples but they are not as amplified. In fact, the smallest

Air Quality Modeling

385

Figure 9.19 O 3 predictions in the rotating/reacting cones test: exact solution (top left), BOT (top right), BOT-M (bottom left), and PPM (bottom right).

RMS error and only a 5% underprediction of the valley may be signs that BOT’s solution for NO, is the best among all the schemes. ASD overpredicts the NO, valley by approximately 4% but there is a slight shape distortion in Fig. 9.17. This distortion may be the reason for an RMS error slightly larger than BOT’s. FCT overpredicts the valley by approximately 10% while the RMS error is smaller than ASD’s. Other schemes perform reasonably, however YAM’S overprediction of the valley by 25% is of concern. SLT overpredicts the valley by 49% and leads to the highest RMS error (7%). The poor performance of many schemes in this case may be due

386

Odman and Russell

Figure 9.20 0, predictions in the rotating/reacting cones test: exact solution (top left), ASD (top right), YAM (bottom left), and FCT (bottom right).

to the extremely small magnitude of NO, concentrations after 48 h. It is important to note that one snapshot at the end of a 48-h simulation does not fully reflect how well the schemes followed NO, as ozone was being formed. The best performance in predicting the 0, peak is displayed by ASD and YAM. The RMS errors for these two schemes are also much smaller than the other schemes. The third best performer appears to be BOT, despite some ring-shaped ripples around the ozone puff. Recall that ASD and BOT also predicted both HC and NO, very well. The predictions for

Air Quality Modeling Table 9.6

387

Rotating/Reacting Cones Test

~~~~~~

NO2

HC Scheme

Peak

(O/O)

0 3

RMS (%)

Valley (%)

RMS (%)

1.91 2.90 3.26 2.44 3.89 2.60 5.00 5.56 2.21

103.8 94.6 108.1 109.7 91.8 114.4 149.0 100.5 125.0

2.59 2.00 2.58 2.22 3.88 2.25 7.10 4.93 2.80

Peak (Oh) RMS (%)

ASD BOT BOT-M FCT HAT PPM S LT SMO YAM

98.7 102.4 98.2 98.9 98.5 96.5 90.9 92.9 98.7

99.8 97.5 89.4 91.1 93.4 87.8 75.3 88.5 99.6

0.51 0.92 1.70 0.96 1.79 1.21 3.20 3.56 0.62

these ozone precursors are equally as important as the predictions of ozone concentrations. YAM’S NO2 predictions at the end of the 48-h simulation were not as good, but (as suggested earlier) this was probably due to the small magnitude of NO2 at the end of the test; the predictions were much better when ozone was being formed. The next four best performers are FCT, BOT-M, HAT, and PPM, respectively. SMO predicts the O 3 peak very well but ripples in precursor fields also lead to large ripples in the O3 field. For this reason, SMO’s RMS error is large. SLT is the worst performer in predicting ozone, as expected based on its poor performance in predicting the precursors. 9.3.3 Results and Discussion

Eight advection schemes (not counting the variations of Bott scheme) were compared using test cases ranging from very simple one-dimensional advection to more robust two-dimensional shear and chemically reactive flows. Most schemes performed consistently in all the tests, but a few failed in some stressful tests. The differences between performance were more pronounced in some tests, and the ranking differed from test to test. However, important properties of the schemes were identified as a result of the diversity of the tests. This preliminary testing of the schemes resulted in the following findings : 0

ASD has very high accuracy except for the smallest (i.e., 2Ax) wavelengths, but even there the accuracy was still better than with most

388

Odman and Russet1

other schemes. ASD is not strictly mass conservative and can cause a slight mass imbalance. It may also lead to mild ripples and overshots (i.e., it is not monotonic), but usually yields a positive definite result. BOT is highly accurate and mass conservative. It may create ripples and overshots, but it is positive definite. The phase-speed errors are usually small and the ripples do not seem to grow under the influence of nonlinear chemistry. BOT-M and BOT2D significantly reduce the accuracy of the original scheme. A mass conservation problem was discovered in the formulation of these schemes. The monotonic flux limitation artificially adds mass downwind from steep negative gradients. This problem was also confirmed by Bott (personal communication, 1995). FCT is fairly accurate and mass conservative. It does not create any ripples or overshots, but does so at the expense of diffusion to the background. HAT has fair accuracy and is mass conservative. However, it may lead to ripples that can grow and cause instabilities, especially near point sources close to the domain boundaries. Its performance is unpredictable under strong shear flow situations. The accuracy may be reduced under constant convergence or divergence, but a convergence followed by divergence, or vice versa, usually leads to high accuracy. P P M also has fair accuracy and is strictly mass conservative. It is monotonic, so it does not lead to ripples or overshots under any circ*mstances. While the diffusion to the backgroucd is very small, peak clipping can be significant. The steepening operator of PPM seems to transform even mild gradients to shock waves. SLT has poor accuracy and severe mass conservation problems. The only positive feature of SLT is that it does not lead to ripples. Though it was not obvious from the test here, it can conceptually withstand shear situations better than the schemes that split horizontal diffusion in two one-dimensional operators. SMO has relatively low accuracy. It is mass conservative but it may lead to ripples in concentration fields. Although the ripples do not cause negative concentrations (i.e., SMO is positive definite), they are associated with relatively large phase-speed errors. Also, in the presence of nonlinear chemistry, the ripples may be amplified. YAM has very high accuracy, even for the shortest wavelengths, and is mass conservative. It can lead to overshots under certain situations. YAM can also create mild ripples but it is positive defi-

Air Quality Modeling Table 9.7

389

Computational Performance

CPU time (Norm = SMO)

ASD

BOT

BOT-M

FCT

HAT

PPM

SLT

SMO

YAM

4.38

0.85

1.23

1.45

1.17

1.24

1.25

1.00

2.25

nite. Its performance in following point-source plumes is much better than that of the other schemes. In addition to the properties discussed above, the computational performance of the schemes was also considered in the ranking. The CPU times for all the tests were averaged and normalized with respect to SMO (Table 9.7). The only scheme that is less CPU intensive than SMO is BOT; other schemes require more C P U time. ASD is the most CPU-intensive scheme as it requires more than four times more CPU than SMO does. However, the C P U time spent during horizontal advection is usually a small fraction of a photochemical model’s total CPU time (most of the C P U time is spent in chemistry). Therefore, even with ASD, an air quality model’s total C P U time would only increase by a fraction (most likely less than 50%). 9.4

EVALUATION OF ADVECTION SCHEMES IN AN AIR QUALITY MODEL AND THE IMPACT ON PREDICTIONS

The above tests show that in the cases studied, some, if not marked, differences are found in the results from the various methods. However, those cases have much more severe gradients than one might expect in an actual application. A logical question then is: will real and significant differences still be found in an actual case? This is tested by implementing some of the methods in an air quality model and applying the model to a typical situation. The application was to the Los Angeles, California, area. A three day smog event, August 27-29, 1987, was simulated. This period was from the Southern California Air Quality Study (SCAQS), a large field program designed to develop the data necessary to conduct more comprehensive air quality modeling. This is one of the SCAQS intensive monitoring periods and the episode has been extensively used for modeling purposes (SCAQMD, 1990; Harley et al., 1993). One of the most widely used air quality models is UAM-IV, and it was used here for the tests. Four schemes were tested : SMO, BOT, YAM, and ASD.

390

Odman and Russet1

It should be noted that there is a major difference between air quality modeling and most other C F D applications as the solution of the advection portion of the problem takes a small portion of the total computational time. For example, in a current application, only about 10% of the total CPU time is spent solving the equations discussed above. Most of the time is spent solving the equations governing the chemical dynamics. Thus, using an advection solver that takes twice as long and is only slightly more accurate may be attractive since the total CPU time increases only about 10%. In light of the findings above, it is thus apparent that ASD is the only method that would appreciably change the required computational time in the applications below.

9.4.1 Tracer Experiments

Simulating a tracer experiment is another test that allows more realistic evaluations (Brost et al., 1988; Odman et al., 1995). All the test cases in section 9.3 are simple enough to have an analytical solution. They do not provide sufficient information about the behavior of the schemes under real atmospheric conditions. In actual simulations, there can be various factors not accounted for in simple test cases (e.g., variable winds, wide range of Courant numbers). Therefore, there is a need to evaluate the performance of the schemes, at least comparatively, in more realistic simulations. To simulate a tracer experiment, all the transport processes of the model must be activated. Evaluating the uncertainty due to the solution of horizontal advection is difficult because of all the other sources of uncertainties involved. In general, uncertainties are introduced by the prognostic or diagnostic wind fields, the computation of eddy diffusivities, and the tracer data itself. However, the test may be viable for a comparison of different schemes assuming that other uncertainties would affect the performance of different schemes in the same way. Odman et al. (1996) simulated a tracer experiment from SCAQS using ASD, BOT, YAM, and SMO schemes. As expected, ASD, BOT, and YAM predicted a less diffused puff with a higher peak concentration than SMO. The trajectories of the plumes were very similar and the peaks were in the same grid cell. YAM predicted the highest peak concentration, as expected based on our experience with the skew advection of a point-source plume test. The second highest peak was predicted by BOT followed by ASD. Recall that the peak ratio for BOT was much higher that ASD, near the source, in the skew advection of a point source plume test. The puff did not travel a long enough distance in this experiment for ASD to show its performance in preserving the peaks farther downwind.

Air Quality Modeling

391

Table 9.8 Performance Evaluation of Ozone Predictions versus Observations for UAM Using the ASD, BOT, YAM, and SMO Advection Schemes

ASD

BOT

-23.6 - 23.6 -3.1 1 22.0 14.4 1.11 47.3 3.56 21.7

- 24.0 - 24.0 6.51 21.8 6.60 0.50 49.8 3.73 24.4

Metric Peak (paired in time, space) ["/O] Peak (paired in space) [Yo] Peak (paired in time) [Yo] Peak (unpaired) [TO] Norma IIzed b Ia s [ O h ] Mean bias [pphm] Normalized error [Yo] Mean error [pphm] Variance [TO]

9.4.2

YAM

SMO

- 25.9

-31.4 -31.4 12.5 27.4 8.43 0.65 50.76 3.80 25.37

- 25.9

0.23 18.0 8.93 0.72 49.3 3.70 23.8

Evaluation of Advection Schemes with Full Chemistry

While the tracer experiment is one indication of how well a solver may work, the true test is one in which the conditions (such as meteorology and emissions) are similar to those in which it will be used to determine control strategies-that is, the set of emissions controls that an area should apply to reach attainment with the ambient air quality standards. Here, we investigate not only how closely the various methods compare in terms of the predictions to the base case (i.e., without emissions reductions) but also if the response to emissions controls vary with the choice of solver. If a bias is found in the response to controls, this would raise questions as to how confidently a method could be used. Table 9.9 Performance Evaluation of Nitrogen Dioxide Predictions versus Observations for UAM Using the ASD, BOT, YAM, and SMO Advection Schemes

Metric Peak (paired in time, (space) [Yo] Peak (paired in space) [YO] Peak (paired in time) [Yo] Peak (unpaired) [Yo] Nor ma I ized bias [YO] Mean bias [pphm] No rma Iized error [%I Mean error [pphm] Variance [O/0]

ASD

BOT

YAM

SMO

-68.1

-61.0

- 65.0

- 63.7

-40.9 -19.1 4.1 1 -2.31 -0.71 40.4 1.94 7.09

-31.8 - 17.1 10.1 - 6.2 - 0.80 38.9 1.92 6.44

- 36.3 - 20.4 8.02 - 5.89 - 0.79 39.0 1.93 6.57

- 22.5 -21.1 21.1 -5.19 - 0.79 39.7 1.94 6.50

392

Odman and Russell

Table 9.10 Correlations of 0, Predictions for Base Case and 50% ROG Reductions

I ndependent (x )

Base Case y=mx+b d e P e n d e n

ASD

m b

R2 BOT

m b

R2 SMO

ASD

YAM

m b R2 m b R2

1,000 0.000 1.000 0.978 - 0.078 0.951 0.997 -0.101 0.951 0.984 - 0.066 0.967

y=mx+b ASD

SMO

m b

1.000 0.000 1.000 0.889 0.378 0.890 0.952 0.145

R2 m b R2

0.898 0.908 0.325 0.922

m b

d e n

ASD

R2 BOT

YAM

SMO

YAM

0.973 0.427 0.951 1.000 0.000 1.000 1.014 0.025 0.987 0.993 0.107 0.989

0.954 0.447 0.951 0.974 0.065

0.982 0.306 0.967 0.996 -- 0.030 0.989 1.011 -- 0.014 0.979 1.000 0.000

0.987 1.000 0.000 1.000 0.968 0.164 0.979

1.000

Independent ( x )

50% ROG

d e P e

BOT

SMO

YAM

1.001 0.268 0.890 1.000 0.000 1.000 1.055 - 0.174 0.980 0.993 0.098 0.979

0.943 0.463 0.898 0 929 0.273 0 980 1.000 0.000

1.015 0.128 0.922 0.986 0.020 0.979 1.043 --0.172 0.966 1.000 0.000 1.000

1.000 0.926 0.352 0.966

Air Quality Modeling

393

Table 9.10 Continued.

Independent (.U)

50% NO, y=mx+h

d e P e n d e n t (4’)

ASD BOT SMO YAM

m h R2 m h R2 rn b R2 m h R2

ASD

BOT

SMO

YAM

1.Ooo O.OO0 1.000 1.003 -0.121 0.966 1.002 -0.061 0.962 1.005 -0.086 0.975

0.962 0.356 0.966 1.ooo O.OO0 1.000 0.996 0.08 1 0.990 0.993 0.094 0.992

0.960 0.324 0.962 0.994 --0.012 0.990 1.Ooo 0.000 1 .ooo 0.987 0.083 0.982

0.970 0.259 0.975 0.999 - 0.04 1 0.992 0.994 0.041 0.982 1.000 0.000 1.000

50% ROG/NO, y=mx+h

d e P e n d e n t (Y)

ASD

SMO

m b R2 m h R2 m

YAM

h R2 m

BOT

Independent (.U)

h R2

ASD

BOT

1.Ooo O.OO0 1.OO0 0.984 -0.029 0.948 0.993 -0.016 0.946 0.988 - 0.007 0.963

0.964 0.360 0.948 1.000 0.000 1.ooo 1.004 0.047 0.987 0.991 0.106 0.990

SMO 0.953 0.360 0.946 0.983 0.033 0.987 1.ooo 0.000 1.000 0.975 0.134 0.978

YAM 0.975 0.244 0.963 0.998 -‘0.036 0.990 1.002 0.006 0.978 1 .000 0.000 1.ooo

Odman and Russell

394

Table 9.11 Correlations of NO, Predictions for Base Case and 50% ROG Reductions Independent ( x )

Base Case y= mx+ b d e P e n d e n

ASD

BOT

SMO

(v)

YAM

m b R2 m b R2 m b R2 m b R2

ASD

BOT

SMO

YAM

1.000 0.000 1.000 1.050 - 0.042 0.963 1.041 - 0.035 0.964 1.036 - 0.039 0.974

0.917 0.066 0.963 1.000 0.000 1.000 0.984 0.012 0.985 0.976 0.010 0.991

0.926 0.059 0.964 1.001 - 0.001

0.940 0.055 0.974 1.015 -.0.003 0.991 0.999 0.008 0.978 1.000 0.000 1.000

Independent ( x )

50% ROG y= mx+ b d e P e n d e n

ASD

BOT

SMO

(v)

YAM

0.985 1.000 0.000 1.ooo 0.979 0.008 0.978

m b R2 m b R2 m b R2 m b R2

ASD

BOT

SMO

1.ooo 0.000 1.000 1.042 - 0.051 0.968 1.036 - 0.044 0.969 1.039 - 0.051 0.975

0.928 0.076 0.968 1.000 0.000

0.936 0.067 0.969 1.002 - 0.003 0.989 1.000 0.000 1 .ooo 0.991 0.003 0.981

1.000 0.987 0.013 0.989 0.989 0.006 0.993

YAM

0.938 0.070 0.975 1.003 0.000 0.993 0.990 0.013 0.981 1.000 0.000 1.000

Air Quality Modeling

395

Table 9.11 Continued. ~~~~

Independent (x)

50% NO, ASD

y=mx+b

d e

e n d e n t

(Y 1

ASD BOT SMO

m b R2 m b R2 m b

R2 YAM

y=mx+b d e

e n

ASD

m b R2

BOT

m b

e n t (Y )

SMO

R2 m b

R2 YAM

m b R2

.ooo 1.ooo 1

0.Ooo

1.084

- 0.026

BOT

SMO

YAM

0.873 0.039 0.947 1.ooo

0.886 0.034 0.95 1 1.006 - 0.002 0.986 1.Ooo

0.9 17 0.028 0.965 1.032 --0.007 0.985 1.012 0.000 0.973 1.Ooo 0.000 1.Ooo

O.OO0 1.Ooo

0.947 1.072 - 0.022 0.95 1 1.053 - 0.020 0.965

0.980 0.006 0.986 0.955 0.01 1 0.985

ASD

BOT

SMO

0.894 0.040 0.960 1.000

0.905 0.03 5 0.960 1.005 -.0.002 0.984 1.Ooo

1.Ooo O.OO0 1.Ooo 1.074

-0.028

0.960 1.060 - 0.023 0.960 1.05 1 -0.023

0.973

0.0oO 1.Ooo 0.979 0.008 0.984 0.967 0.009 0.988

O.OO0 1.OO0

0.96 1 0.008 0.973

O.OO0 1.Ooo

0.972 0.007 0.974

YAM 0.925 0.03 1 0.973 1.022

- 0.005

0.988 1.002 0.002 0.9 74 1.OO0 0.000 1.Ooo

Odman and Russell

396 Table 9.12 Strategies

SMO ASD BOT YAM

Ozone Peak Predictions and Relative Reductions for Three Control

50%

Base-50%

50%

Base

ROG

NO,

36.9 35.4 35.3 34.2

21.5 20.3 19.1 18.1

15.4 15.1 16.2 16.1

39.5 36.5 38.4 38.4

Base-5O0/o NO, - 2.6 -1.1 -3.1 - 4.2

50% ROG/NO,

Base-50% ROG/NO,

29.1 27.8 27.5 28.0

7.8 7.6 7.8 6.2

The results from applying the model for both the base case inventory and the control strategy inventories using the four advection schemes are given in Tables 9.8-9.12. As can be seen, the three more accurate methods (BOT, YAM, and ASD) compare well, while SMO shows some deviation from the rest. The response to controls was also very similar (Tables 9.10 and 9.11). The amount of reduction each advection solver predicts is approximately the same for the 50% ROG and the 50% ROG/NO, cases. There is more variability between the solvers for the 50% NO, case, but the numbers are all fairly small. Thus, the solvers may not predict the same absolute peak ozone, but the reductions for each control strategy are similar in most cases (Table 9.12). This test indicates that the choice of the solvers does not bias the choice of a control strategy. In a more recent study (Krishnakumar et al., 1998), a case was found where the differences were more marked. In this case, UAM-IV was applied to the Pittsburgh, Pennsylvania region using the S M O and BOT routines. They found that the S M O scheme led to consistently higher ozone levels (a few percent) and that there were a few erroneously high predictions. Further, the peak ozone levels (e.g., those erroneously high predictions from using SMO) did not respond to controls in a similar fashion between the use of S M O and BOT. This is particularly of concern given how air quality modeling is used in a regulatory framework. Usually, greatest (if not complete) importance is placed on the peak ozone predicted anywhere in the domain. The results of that study raise questions as to the confidence one has in determining control strategies using the SMO. 9.5

SUMMARY

One of the most important components of an air quality model is the scheme used to integrate the advection portion of the governing equation. Horizontal transport in the atmosphere is dominated by advection, while vertical transport is dominated by diffusion. Thus, the horizontal transport

Air Quality Modeling

397

solver must be able to handle the hyperbolic nature of the problem without introducing significant diffusion or dispersion. In the presence of the nonlinear chemistry, such errors can significantly degrade model results, particularly in test cases. A number of methods have evolved, and the current methods, including finite element, finite volume, pseudospectral, and advanced finite difference methods, give similar results in most test cases. One of the critical components of those methods is how to filter out the dispersive waves that lead to artificial minima and maxima (including negative concentrations) while not degrading the solution by adding excessive artificial diffusion. There has been relatively little intercomparison of the methods in actual applications, i.e., comparison of how the methods work when applied to atmospheric modeling of specific events. Those few studies that have investigated the problem find that while the results are similar, there can be some problems. For example, the response to emissions controls, as well as peak predictions (which are of extreme importance in air quality modeling), can be in error if less accurate methods are used. This is particularly true with older methods that have been found to be diffusive. However, when a variety of methods that have been shown to perform well across a suite of test cases is used, the results in actual applications of those methods are similar. REFERENCES

Anderson DA, Tannehill JC, Pletcher RH. Computational Fluid Mechanics and Heat Transfer. Washington, DC: Hemisphere, 1984. ARB. Technical Guidance Document: Photochemical Air Quality Modeling. Modeling and Meteorology Branch, Technical Support Division, California Air Resources Board, Sacramento, CA, 1990. Aron JD, Vicente F-G. The random choice method in the numerical solution of the atmospheric transport equation. Environ Software 9 : 23-3 1, 1994. Bell J, Berger M, Saltzman J, Welcome M.Three-dimensional adaptive mesh refinements for hyperbolic conservation laws. SIAM J Sci Comput 15(1): 127-138, 1994. Book DL, Boris JP, Hain K. Flux-corrected transport. 11. Generalizations of the method. J Comp Phys 18: 248-283,1975. Boris JP, Book DL. Flux-corrected transport. 1. SHASTA, A fluid transport algorithm that works. J Comp Phys 11: 38-69, 1973. Boris JP, Book DL. Flux-corrected transport. 111. Minimal error FCT algorithms. J Comp Phys 20: 397-431, 1976. Bott A. A positive definite advection scheme obtained by nonlinear renormalization of the advective fluxes. Mon Wea Rev 117: 1006-1015, 1989a. Bott A. Reply. Mon Wea Rev 117: 2633-2636,1989b.

Odman and Russell

398

Bott A. The monotone area-preserving flux-form advection algorithm: reducing the time-splitting error in two-dimensional flow fields. Mon Wea Rev 121 : 26372641, 1993.

Bott A. Monotone flux limitation in the area-preserving flux-form advection algorithm. Mon Wea Rev 120: 2592-2602,1992. Brooks AN, Hughes TJR. Streamline upwind/Petrov-Galerkin formulations for convection dominated flows with particular emphasis on the incompressible Navier-Stokes equations. Comp Meth Appl Mech Eng 32: 199-259, 1982. Brost RA, Haagenson PL, Kuo Y-H. The effect of diffusion on tracer puffs simulated by a regional scale Eulerian model. J Geophys Res 93: 2389-2404, 1988. Carpenter RL, Droegemeier KK, Woodward PR, Hane CE. Application of the piecewise parabolic method (PPM) to meteorological modeling. Mon Wea Rev 118: 586-612, 1990. Childs PN, Morton KW. Characteristic Galerkin methods for scalar conservation laws in one dimension. SIAM J Numer Anal 27: 553-594,1990. Chlond A. Locally modified version of Bott’s advection scheme. Mon Wea Rev 122: 11 1-125, 1994.

Chock DP. A comparison of numerical methods for solving the advection equation. 11. Atmos Environ 19(4): 571-586, 1985. Chock DP. A comparison of numerical methods for solving the advection equation. 111. Atmos Environ 25A(5/6): 853-871, 1991. Chock DP, Dunker AM. A comparison of numerical methods for solving the advection equation. Atmos Environ 17( 1): 11-24, 1983. Chock DP, Winkler SL. A particle grid air quality modeling approach. 2. Coupling with chemistry. J Geophys Res 99: 1033-1042, 1994a. Chock DP, Winkler SL. A comparison of advection algorithms coupled with chemistry. Atmos Environ 28: 2659-2675, 1994b. Chorin AJ. Random choice solution of hyperbolic systems. J Comp Phys 22: 517 533, 1976.

Colella P, Woodward PR. The piecewise parabolic method (PPM) for gasdynamical simulations. J Comp Phys 54: 174-201, 1984. Concus P, Proskuroski W. Numerical solution of a nonlinear hyperbolic equation by the random choice method. J Comp Phys 30: 153-166, 1979. Crowley WP. Numerical advection experiments. Mon Wea Rev 96: 1-1 1, 1968. Dabdub D, Seinfeld JH. Numerical advective schemes used in air quality models-sequential and parallel implementation. Atmos Environ 28(20): 3369-3385. 1994.

Donea J. A Taylor-Galerkin method for convective transport problems. Int J Num Meth Eng 20: 101-1 19, 1984. Easter RC. Two modified versions of Bott’s positive-definite numerical advection scheme. Mon Wea Rev 121: 297-304, 1993. Emde KVD. Solving conservation laws with parabolic and cubic splines. Mon Wea Rev 120: 482-492, 1992. Flatoy F. Balanced wind in advanced advection schemes when species with long lifetimes are transported. Atmos Environ 27A( 12): 1809-1 819, 1993.

Air Quality Modeling

399

Fogelson AL. Particle method solution of two-dimensional convection-diffusion equations. J Comput Phys 100: 1-16, 1992. Forester CK. Higher order monotonic convective difference schemes. J Comp Phys 23: 1-22,1977. Gazdag J. Numerical convective schemes based on accurate computation of space derivatives. J Comp Phys 13: 100-113, 1973. Glimm J. Solutions in the large for nonlinear hyperbolic systems of equations. Comm Pure Appl Math 18: 697-715, 1965. Godunov SK. A difference scheme for numerical computation of discontinuous solution of hydrodynamic equations. Math Sb 47: 271-306, 1959 (in Russian). Harley RA, Russell AG, McRae GJ, Cass GR, Seinfeld JH. Photochemical air quality modeling of the Southern California air quality study. Environ Sci Tech 27: 378-388, 1993. Hirt CW. Heuristic stability theory for finite-difference equations. J Comp Phys 2 : 339-355, 1968. Hov 0, Zlatev Z, Berkowicz R, Eliassen A, Prahm LP. Comparison of numerical techniques for use in air pollution models with nonlinear chemical reactions. Atmos Environ 23: 967-983,1989. Hughes TJR, Brooks AN. A multidimensional upwind scheme with no crosswind diffusion. In: Finite Element Methods for Convection Dominated Flows. ASME, 1979. Hughes TJR. Recent progress in the development and understanding of SUPG methods with special reference to compressible Euler and Navier-Stokes equations. In: Finite Elements in Fluids. Vol 7. New York: Wiley, 1987, pp 273-287. Kelly DW, Nakazawa S , Zienkiewicz OC, Heinrich JC. A note on upwinding and anisotropic balancing dissipation in finite element approximations to convective diffusion problems. Int J Num Meth Eng 15: 1705-1711, 1980. Kreiss H - 0 . Some remarks about computational fluid dynamics. In: Hussaini MY, Kumar A, Salas MD, eds. Algorithmic Trends in Computational Fluid Dynamics. New York : Springer-Verlag, 1993. Krishnakumar V, Wilkinson J, Russell A. Air quality modeling of Southwestern Pennsylvania: Errors in peak ozone predictions by the horizontal advection equation solver in the urban airshed model. J Air and Waste Manage Assoc (submitted). Lohner R, Morgan K, Peraire J, Vahdati M. Finite element flux-corrected transport (FEM-FCT) for the Euler and Navier-Stokes equations. In: Finite Elements in Fluids. Vol7. New York: Wiley, 1987, pp 105-121. Marchuk GI. Methods of Numerical Mathematics. Springer-Verlag, 1975. McRae GJ, Goodin WR, Seinfeld JH. Numerical solutions of the atmospheric diffusion equation for chemically reacting flows. J Comp Phys 45: 1-42, 1982. Morris RE, Myers TC. User’s guide for the urban airshed model. Vol I. User’s manual for UAM (CB-IV). Research Triangle Park, NC: Environmental Protection Agency, 1990. Morton K W, Stokes A. Generalized Galerkin methods for hyperbolic problems.

Odman and Russell

400

Proceedings of the Conference on Mathematics of Finite Elements and Applications IV. New York: Academic Press, 1982, pp 421-431. M uller R. The performance of classical versus modern finite-volume advection schemes for atmospheric modeling in a one-dimensional test-bed. Mon Wea Rev 120: 1407-1415, 1992. NRC. Rethinking the Ozone Problem in Urban and Regional Air Pollution. Washington, DC: National Academy Press, 1992. Odman MT. Multiscale modeling of long range transport and fate of pollutants emitted from urban areas. PhD thesis, Carnegie Mellon University, Pittsburgh, PA, 1992. Odman MT, Russell AG. A multiscale finite element pollutant transport scheme for urban and regional modeling. Atmos Environ 25A: 2385-2394, 1991b. Odman MT, Russell AG. Multiscale modeling of pollutant transport and chemistry. J Geophys Res 96: 7363-7370, 1991a. Odman MT, Russell AG. A nonlinear filtering algorithm for multi-dimensional finite element pollutant advection schemes. Atmos Environ 27A : 793-799, 1993.

Odman MT, Wilkinson JG, McNair LA, Russell AG, Ingram CL, Houyoux MR. Horizontal advection solver uncertainty in the urban airshed model. Final report, contract no. 93-722, California Air Resources Board, Sacramento, CA, 1996.

Odman MT, Xiu A, Byun DW. Evaluating advection schemes for use in the next generation of air quality modeling systems. In: Ranzieri AJ, Solomon PA, eds. Regional Photochemical Measurement and Modeling Studies. Pittsburgh, PA: Air & Waste Management Association, 1995, pp 1386 1401. Olim M. A truly noninterpolating semi-Lagrangian Lax-Wendroff method. J Comp Phys 112: 253-266, 1994. OTA (Ofice of Technology Assessment, U.S. Congress). Catching our breath-next steps for reducing urban ozone. OTA-0-412, Congressional Board of the lOlst Congress. Washington, DC: U.S. Government Printing Office, 1989. Pai P, Karamchandani P, Venkatram A. Performance of flux conserving and semiLagrangian advection schemes in simulating a photochemical episode. 2 1 st NATO/CCMS International Technical Meeting on Air Pollution Modeling and Its Application, Baltimore, M D, November 6- 10, 1995. Parrott AK, Christie MA. FCT applied to the 2-D finite element solution of tracer transport by single phase flow in a porous medium. In: Numerical Methods for Fluid Dynamics. 11. New York: Oxford University Press, 1986, pp 609619.

Pepper DW, Kern CD, Long PE. Modeling the dispersion of atmospheric pollution using cubic splines and chapeau functions. Atmos Environ 13 : 223-237, 1979. Pielke RA. Mesoscale Meteorological Modeling. New York : Academic Press, 1984. Purnell DK. Solution of the advective equation by upstream interpolation with a cubic spline. Mon Wea Rev 104: 42-48, 1976. Rasch PJ, Wiliamson DL. O n shape-preserving interpolation and semi-Lagrangian transport. SIAM J Sci Stat Comput 1l(4): 656-687, 1990.

Air Quality Modeling

40 1

Raymond WH, Garder A. Selective damping in a Galerkin method for solving wave problems and variable grids. Mon Wea Rev 104: 1583-1590, 1976. Roache PJ. A pseudo-spectral FFT technique for non-periodic problems. J Comp Phys 27 : 204-220, 1978. Roberts PT, Main HH. Ozone and particulate matter case study analysis for the Southern California air quality study. Prepared for the California Air Resources Board by Sonoma Technology, Inc., contract A932-050, 1993. Rood RB. Numerical advection algorithms and their role in atmospheric transport and chemistry models. Rev Geophysics 25( 1): 71-100, 1987. Scheffe RD, Morris RE. A review of the development and application of the urban airshed model., Atmos Environ 27B: 23-39, 1993. SCAQMD. Air quality management plan, technical report V-B, ozone modelingperformance evaluation. South Coast Air Quality Management District, 1990. Seibert P, Morariu B. Improvements of upstream, semi-Lagrangian numerical advection schemes. J Appl Meteor 30: 1 17-1 25, 1991. Seinfeld JH. Atmospheric Chemistry and Physics of Air Pollution. New York: John Wiley & Sons, 1985. Seinfeld JH. Ozone air quality models: a critical review. JAPCA 38(5): 616-645, 1988. Smolarkiewicz PK. A fully multidimensional positive definite advection transport algorithm with small implicit diffusion. J Comp Phys 54: 325-362, 1984. Smolarkiewicz PK. A simple positive definite advection transport scheme with small implicit diffusion. Mon Wea Rev 11 1 : 479-486, 1983. Smolarkiewicz PK, Grabowski WW. The multidimensional positive definitive advection algorithm: nonoscillatory option. J Comp Phys 84: 355-375, 1990. Smolarkiewicz PK, Grell GA. A class of monotone interpolation schemes. J. Comp Phys 101 : 431-440, 1992. Strand A, Hov 0. A two-dimensional zonally averaged transport model including convective motions and a new strategy for the numerical solution. J Geophys Res 98: 9023-9037, 1993. Staniforth A, C6te J, Pudykiewicz J. Comments on Smolarkiewicz’s deformational flow. Mon Wea Rew 115: 894-900,1987. Tesche TW, et al. Improvement of procedures for evaluating photochemical models. Radian Corporation, Report to the California Air Resources Board, 1990. Tezduyar TE, Park YJ, Deans HA, Finite element procedures for time-dependent convection-diffusion-reaction systems. In : Finite Elements in Fluids. Vol 7. New York: Wiley, 1987, pp 25-45. Tolstykh MA. Application of fifth-order compact upwind differencing to moisture transport equation in atmosphere. J Comp Phys 112: 394-403, 1994. Tremback GJ, Powell J, Cotton WR, Pielke RA. The forward-in-time upstream advection scheme: extension to higher orders. Mon Wea Rev 1 1 5: 540-555, 1987. van Leer B. Towards the ultimate conservative difference scheme: IV. A new approach of numerical convection. J Comp Phys 23: 276-299, 1977. Wengle H, Seinfeld JH. Pseudospectral solution of atmospheric diffusion problems.

Odman and Russell

402

J Comp Phys 26: 87-106, 1978. Westerink JJ, Shea D. Consistent higher degree Petrov-Galerkin methods for the solution of the transient convective-diffusion equation. Int J Num Meth Eng 28: 1077-1 101,1989.

Williamson DL, Rasch PJ, Two-dimensional semi-Lagrangian transport with shapepreserving interpolation. Mon Wea Rev 117: 102-1 17, 1989. Williamson DL, Olson JG. Climate simulations with a semi-Lagrangian version of the NCAR community climate model. Mon Wea Rev 122: 1594-1610, 1994. Yamartino RJ, Scire JS, Carmichael GR, Chang YS. The CALGRID mesoscale photochemical grid model. I. Model formulation. Atmos Environ 26A: 14931512,1992.

Yamartino RJ. Nonnegative, conserved scalar transport using grid-cell-centered, spectrally constrained Blackman cubics for applications on a variablethickness mesh. Mon Wea Rev 121: 753-763, 1993. Yanenko YN. The Method of Fractional Steps. New York: Springer, 1971. Y U C, Heinrich JC. Petrov-Galerkin method for multidimensional, time-dependent. convective-diffusion equations. Int J Num Meth Eng 24: 2201-2215, 1987. Yu C-C, Heinrich JC, Petrov-Galerkin methods for the time-dependent convective transport equation. Int J Num Meth Eng 23: 883-901, 1986. Zalesak ST. Fully multidimensional flux-corrected transport algorithms for fluids. J Comp Phys 31 : 335-362, 1979.

Appendix Governing Equations in Various Systems Vijay K. Garg A Y T CorporationlNASA Lewis Reseach Center. Cleoeland. Ohio

. . . . . . . A.2 Governing Equations in Cylindrical Coordinates . . . . . . . A.3 Governing Equations in Spherical Coordinates . . . . . . . A.4 Orthogonal Curvilinear Coordinates and Vector Field Equations . A.4.1 Vector Field Functions . . . . . . . . . . . . . A.5 Metric Coefficients in Various Orthogonal Coordinate Systems . . A S .1 Circular Cylinder Coordinates . . . . . . . . . . . A.5.2 Elliptic Cylinder Coordinates . . . . . . . . . . . A.5.3 Bipolar Cylinder Coordinates . . . . . . . . . . . A.5.4 Parabolic Cylinder Coordinates . . . . . . . . . . A.5.5 Spherical Coordinates . . . . . . . . . . . . . . A.5.6 Prolate Spheroidal Coordinates . . . . . . . . . . A.5.7 Oblate Spheroidal Coordinates . . . . . . . . . . . A.l Governing Equations in Cartesian Coordinates

A.l

. . . . .

403 405 407 410 412

. . 415

. . 415 . . 415 . . 415

. . 416

. . 416 . . 416 . . 416

GOVERNING EQUATIONS IN CARTESIAN COORDINATES

Coordinates x. y. z ; velocity components in these directions

U. U. w .

The 403

404

equation of continuity for a compressible flow, c:p (7t

(pv)= 0

can be written in Cartesian coordinates as

+-+,-2Piit

?(P4 (l.x

8 ( P U ) , d(PW) - 0 ay

For an incompressible flow, Eq. (A.2) reduces to

-?U+ - +?U- = oaw c:x

(74,

(A.3)

8.2

The momentum equations for a uiscous, compressible Juid are

Du (A.4a1

(A.4b)

+ c7 dx [ P ( Z +

a)]

+ ;

[P(E +

E)]

(A.4c)

where

_D --_Z Dt at

+ v * v = - + iiu - + v ii- + w - a at ax ay

a aZ

and V Vis defined by Eq. (A.3). Equations (A.4) for a uiscous, incompressible Juid reduce to

Du Dt

(A.5a) (A.5b)

D U’ Dt

a2w

aZW

azw

(A.5c)

Governing Equations

405

If we let p = 0 in Eqs. (A.4) we obtain the momentum equations for an inviscid fluid. The energy equation in Cartesian coordinates has the form D ( ~ , T ) DP a - -- - ( k Dt Dt ax

where =

")ax + a( k "> +2 ay i?z ay

"> sz

+ CD

(A.6)

(g)2 (g aw(5 aui.;)-p* +

+ j1

+ ao

+&+-J+&+z) 1 av aw

y)']

For a viscous, incompressible fluid the stress functions are ox,. = 2P

-- P,

tJyy

= 2p - - p ,

6,, =

aw au

azx= p(- az

A.2

aw + -) ax

(A.7a)

-p

(A.7b) (A.7c)

= axz

GOVERNING EQUATIONS IN CYLINDRICAL COORDl NATES

Coordinates r, 8, z ; velocity components in these directions are V , , b, and V ,. The relationship between the cylindrical and Cartesian coordinates is illustrated in Fig. A.l. For a compressible fluid the continuity equation becomes aP -+at

4 P K ) +-1 4 P K ) I d(PK/,) ( P V - 0

(A.8)

For an incompressiblefluid Eq. (A.8) reduces to

-av, + - - +l -d+v-, = oav, v, ar

r 88

(A.9)

The momentum equations for a compressible fluid with constant riscositj, are

DV, V i Dt

(A.lOa)

406

Garg

x = r cos 8 y = r sin 8

z=z

I vi Figure A.l

Relationship between Cartesian and cylindrical coordinates.

ov,= g z - - Dt

_~a--_ Dt

[

ap+v V2v,+--(V* 3l a aZ

P az

a v,a +v*-+--+v,ar

r 80

(A.lOb) (A.lOc)

a aZ

For incompressiblepows Eqs. (A.lO) get simplified since V V = 0. If we let v = 0 in Eqs. (A.lO) we obtain the momentum equations for an inuiscidjuid. The energy equation in cylindrical coordinates becomes

(A. 11)

where

Governing Equations

407

The stress functions for a viscous, incompressible Juid are

(2+ T)

o,, = p -

A.3

- = or,

(A. 12)

GOVERNING EQUATIONS IN SPHERICAL COORDINATES

Coordinates O, 4, 8 ; velocity components in these directions are r/W, V4, and I/e . The relationship between the Cartesian, cylindrical, and spherical coordinates is illustrated in Figs. A.2 and A.3. For a compressiblepuid the continuity equation becomes aP 1 a(O2PV") 1 d(Pv,) -++-~ s1i 4n a(oV4&psin 4) +--=0 at o2 aO o sin 4 ae

For an incompressiblefluid Eq. (A.13) reduces to

Figure A.2

Relationship between Cartesian and spherical coordinates.

(A.13)

Garg

408 U

r = w sin a~

e=e

= w cos al

Figure A.3

Relationship between cylindrical and spherical coordinates.

(A.14) The momentum equations for a compressible .fluid with constant ciscosity are DV, -

V:+ V i

=Y,---

10 P aw

v,,

CO2

V2Yo-2-----

av, 84

2v, cot CO2

w 2 sin

"c7eI (A. 15a)

--.+.I(_+-i?2v, 2 i?vw cot 4 %

3 [ co2 l

(?#2 72v

+ w 2 sin

?+ ?w

(-(72

4 ?4 iv

w )

cot

c#)

+ w2 # j

- V@)]

(A.15 b)

409

Governing Equations

V+&cot

DV, V V -+++ o Dt

+--

v 1 3 o sin

sin2

ge - p o sin

2 sin

av,

4 -+ 80

2 cos 4 sin2 4 20

(A.15~)

where

For the incompressiblejuid Eqs. (A.15) reduce to

v$ + v; w

1 aP =sw---

vmv+ v;

cot

=go---

2 D V , r/e V, Dt o

V+Ve cot w

av,

1 aP PO

a+ 1

se - p o sin

2 cos

4 2b

(A. 16b)

(A. 16c)

If we let v = 0 in Eqs. (A.15) we obtain the momentum equations for an inviscid fluid. The energy equation in spherical coordinates becomes D(c,T) Dt

Dp Dt

w2 d o

1 w2 sin 1

4 &i

Garg

410

where

+(--

1 cu sin

av, vw v, cot 4

4 80

+-+ o

+ -1 {-sin 4 a (") 2

84 sin (b

+1 av w sin (b 80

The stress functions for a uz'scous, incompressiblefluid are QWW

aaw

= 2P VW - p ,

CT,,=2p

;:(

--p+-

- p

(A. 18)

A.4

ORTHOGONAL CURVILINEAR COORDINATES AND VECTOR FIELD EQUATIONS

Let xl, x 2 , and x 3 be a set of orthoyonal curvilinear coordinates; i.e., they are mutually at right angles to each other at their point of intersection. The relation between Cartesian coordinates (x, y, z) and curvilinear coordinates may be written as (A. 19)

41 1

Governing Equations

In regions where the Jacobian

(A.20)

aZ aZ az, aZ2 az3 we can invert the transformation (A.19) to get (A.21) If J = 0 this implies that xl, x 2 , and x3 are not independent, but are connected by some functional relationship of the form Curvilinear coordinates may be interpreted geometrically as in Fig. A.4. The element ds of the curve passing through the points P and Q in Cartesian coordinates is ds = i d x

+ k dz,

i,j , k are unit vectors

(A.22)

Equation (A.22) may be written = (dx)2

+ (dy)2+ (d2)2

Additionally

(A.23)

( d ~=) (~h , d ~ , ) (h2 ~ d ~ , ) ~(h3 d ~ , ) ~

(A.24)

where

(A.25)

h,, h 2 , and h3 are termed metric coefficients or scale factors. It should be noted that unit vectors in the curvilinear system are also transformed and

Garg

412

Figure A.4

Curvilinear coordinates.

change with the system. In order to express the governing equations in any ort hogonal coordinate system, vector field functions involving the metric coefficients in curvilinear coordinates are needed. A.4.1

Vector Field Functions

i,, i,, i, are unit vectors in the directions of x,, x2, and x3 positive. In the

following, 4 is a scalar function, p, q are arbitrary vectors with components p , , p 2 , p , , q , , q 2 , q3 in the xl, x 2 , x, directions. The “directional sense” is the conventional right-handed system. The following relations hold, the functions being invariant; i.e., they do not depend on any particular coordinate system. Grad

1 &p 4 = V 4 = - -il+-- 1 84 i,

h , ?xl

h , C?x2

1 84 +-i, h3 d x ,

(A.26)

(A.27)

Curl p

V xp

= -{

hlh2 h ,

h , [ Y --

413

Governing Equations

hlil

a -

ax, hlP1

v2p

h2i2

h3i3

a -

2 2x3

ax2 h2P2

(A.28)

h3P3

V(V ' p ) - v x (V x p )

Note the distinction between the vector operator V2 and the scalar (div grad). (a(h242) 3(hiqi)) 2x1 8x2

(A.3 1)

414

Garg

The components of the stress tensor can be expressed as

(A.32)

where the expressions for the rates of strain are U, ah, +--+--

u3 dh, hlh, d x ,

h l h , ax,

):( a ($ a ):(

8x1 8x2 8x3

ax,(h,) h2 a ax,(h,) h3 a ax,(G)=

= ex2x1

(A.33)

= ex3x2

ex1x3

and U 3 are the velocity components in the x l , x 2 , and x3 directions, respectively. The components of V sijare U17 U,,

415

Governing Equations

(A.34)

The dissipation function becomes r

(A.35)

The above relations can be used to derive the governing equations in any orthogonal curvilinear coordinate system.

A.5

METRIC COEFFICIENTS IN VARIOUS ORTHOGONAL COORDINATE SYSTEMS

A51

Circular Cylinder Coordinates

dx, = dr h, = 1 x = Y cos

A.5.2

dx, = do h2 = Y y = r sin 8

dx, = dz h3 = 1 z=z

Elliptic Cylinder Coordinates

d x , = dq dx2 = d r dx3 = dz h , = h2 = c(sinh2 sin2 q)lI2 h, = 1 x = c cosh ( cos q y = c sinh ( sin q c = positive constant

A53

dx,

Bipolar Cylinder Coordinates = dq

h, = h , =

dx2 = d r C

cosh q - COS

dx, = dz

h, = 1

z =z

Garg

416

c sinh q cosh v - COS 5 c = positive constant =

A.5.4

z=z

d ~ =, d5 d.u3 = dz h, = 1 q2)li2

h , = h2 = 2 c ( t 2 s = c ( 5 2 - q2) c =

2’ = 2c5q

positive constant

Spherical Coordinates

A.5.5

dY1 =

= c‘

sin $ cos 8

d x , = dd,

h2 = w y = w sin d, sin 8

dx, = de

h , = w sin d, z = w cos d,

=dt

dx, = dq dx, = dd, h , = c(sinh2 5 + sin2q)li2 h , = c sinh 5 sin q sinh { sin q cos 4 z = c cosh 2’ = c sinh 5 sin q sin $

d.Y,

Prolate Spheroidal Coordinates

A.5.6

A57

c sin 5 cosh q - COS 5

Parabolic Cylinder Coordinates

d . ~= , dq

0 blate Spheroidal Coordinates

dx, = d t d x 2 = dq dx, = dd, h , = h, = c(sinh2 5 + cos’ q)1’2 h, = c cosh 5 sin q s = c cosh 5 sin q cos 4 y = c cosh 5 sin q sin d, z = c sinh 5 cos q

5 cos q

Index

Accuracy : first-order, 39 second-order, 39 Adiabatic flow, 7 Advection equation, 343, 350 advective form, 343 flux form, 343 Advection schemes, 343-363 classification of, 345 evaluation of, 363-396 finite difference schemes, 346, 352 finite element schemes, 348, 358 chapeau function scheme, 358-359,363,388 Petrov-Galerkin method, 349 st rearnline upwind/Petrov-Galerkin (SU/PG) method, 349 Taylor-Galerkin method, 349 finite volume schemes, 346, 351-353, 357

[Advection schemes] Bott's scheme, 347, 354-356, 363, 388 piecewise parabolic method (PPM), 347,353-354,363,388 Yamartino's scheme, 347, 363, 388 flux-corrected transport (FCT) scheme, 348,357,363,388 performance measures of: average absolute error, 366 background-to-peak ratio, 364 distribution ratio, 364 mass ratio, 364 peak ratio, 364 root mean square (RMS) error, 367 properties of, 344-345 pseudospectral scheme, 350,359-360 accurate space derivative (ASD) scheme, 350,359-360, 363,387 417

418

[Advection schemes] semi-Lagrangian schemes, 348, 361, 363,388 semi-Lagrangian transport (SLT) scheme, 36 1-363 Smolarkiewicz scheme, 352-3 5 3, 363,388 Tremback’s scheme, 355 Advection tests, 363 advection of 1-D pulses, 367 advection of point-source plume, 373 rotating cone test, 370 rota tinglreacting cones, 380 shear flow tests, 377 tracer experiments, 390 Air pollution problems: acid deposition, 336 global climate warming, 336 smog, 336,389 stratospheric ozone depletion, 336 Air quality model (Eulerian), 337, 341, 389,396 Algebraic equations, 37 Alternating direction implicit (ADI) method, 267 Amplification factor, 47 Anisotropic materials, 28 1 Annealing, 294 Anthropogenic source, 336 Assembly, 53 Atmospheric diffusion equation (ADE), 339 Average-passage approach, 238 Average period approach, 247 Baldwin-Lomax algebraic model, 84-85 Barotropic fluid, 18 Bernoulli equation, 18 compressible, 18 Biogenic source, 336 Block-merging algorithm, 138 Body force, 3 Boundary conditions, 7, 184,212 Boundary element method (BEM), 282

Index

Boundary layer equations, 11 Boussinesq assumption, 80 Bulk viscosity, 4 Buoyancy effects, 263 Cauchy-Riemann conditions, 124 Cebeci-Smith algebraic model, 84 CFL condition, 48 Chien k-E model, 89-90 Combined cycle engines, 205-206 Compressibility, 181 Compressibility effects, 91-92,96, 106 Compressible fluid, 2 Computer hardware, 2 12-2 13 Condensation, 56 Conjugate transport, 290 Continuity equation, 2 Convergence, 48, 188 Courant number, 47 Creeping flow, 273,276 Damping functions, 83, 89 Degrees of freedom, 54 Delaunay triangulation, 159 Design process, 246 Difference : first-order backward, 39 first-order forward, 39 one-sided, 42 second-order central, 40 Direct numerical simulation (DNS), 82,99 Dirichlet polygon, 159 Dirichlet tessellation, 158-1 59 Dissipation function, 6 Distortion, 198-99 Documentation, 216 Duct flows, 179-219 Dynamic viscosity, 4 Ease of use, 213-214 Eckert number, 13,273 Ecological damage, 336 Eddy diffusivity (thermal)models, 102-104 one-equation models, 103

Index

[Eddy diffusivity (thermal) models] two-equation models, 103-104 Eddy viscosity models, 83-92 algebraic models, 83-85 one-equation models, 85-86 two-equation models, 86-92 Electronic equipment : thermal analysis of, 305-306 thermal control of, 302-333 active, 303 passive, 302-333 thermal design challenges, 305 Electronics packaging : direct-chip-attach scheme, 305 hierarchy of, 302 structural levels, 302-303 board, 302-303 chip or component, 302 system, 302-303 thermal issues, 303-305 Element, 54 compatible or conforming, 54 complete, 54 isoparametric, 54 Elliptic equations, 16, 30 Emission controls, 336, 391, 397 Energy equation, 6 Enthalpy method, 265-266 Equation of state, 8 Equidistribution principle, 148 Error : discretization, 45 round-off, 45 truncation, 43 Error of approximation, 45 Error of calculation, 45 Euler equation, 6, 16 Eulerian approach, 3 Eulerian model, 340 Euler-Lagrange equation, 149 False transients, 267 Fast Fourier transform (FFT), 360 Favre averaging, 78-80 Fickian diffusion, 338 Finite difference method (FDM), 37

419

[Finite difference method (FDM)] Crank-Nicolson, 43 explicit, 43 implicit, 43 Finite element method (FEM), 49 direct approach, 49 energy balance approach, 50 variational approach, 49 weighted residual formulation, 49 Finite volume method (FVM), 56 cell-centered, 62 cells, 57 cell vertex, 63 Flux transport (thermal)models, 104-106 Free surfaces, 270 Functional, 49 Galerkin method, 52 Gaussian elimination, 43 Generalized coordinates, 25 Generalized gradient diffusion hypothesis (GGDH), 101-102 Geometric conservation law, 25 Geometry definition, 185 Gibbs phenomenon, 345 Glass, 280 Governing equations : in Cartesian coordinates, 403-405 in cylindrical coordinates, 405-407 in generalized coordinates, 25-29 mathematical properties of, 29-30 nondimensional form of, 10-1 1 in spherical coordinates, 407-410 vector form of, 9 Grashof number, 273 Grid : adaptive, 119 boundary-fitted, 134 chimera, 136 coloring, 170 conformal, 122 multiblock, 136-140 orthogonal, 122 skewness, 1 19 structured, 49, 1 18

420

[Grid] unstructured, 49, 118 Grid density, 186-1 87, 204-205 Grid generation: adaptive, 147-157 algebraic, 140- 147 elliptic, 133- 136 orthogonal, 130-133 structured, 123-147 unstructured, 157-171 Grid generation code: GRAPE, 135 GRIDGEN, 138 GridPro/az3000, 137 JERRY & T O M , 136 STACK3D, 136 TOMCAT, 136 WESCOR, 135 Grid points, 36 Grid quality, 186-187 Green’s function method, 37 High-frequency noise, 36 1 Hot streak, 232 Hyperbolic equations, 29 Incompressible fluid, 3 Indexing of stators, 235 Inlet flows, 179-219 Interface minimization technique, 171 Inviscid fluid. 5 Jacobian, 27 Jones-Launder k-E model, 87 k-E models, 86-91 k - o model, 87 K-theory, 338

Lagrangian approach, 3 Lagrangian model, 340 Lagrangian trajectory, 340 Lam-Bremhorst k-E model, 89-90 Landau transformation, 287 Laplace equation, 20 Large eddy simulation (LES), 83

Index

Launder-Reece-Rodi stress transport model, 93 Launder-Sharma k-E model, 87,89-90 Lax’s equivalence theorem, 48 Liquid immersion cooling, 306-328 dielectric liquid, 306 fluorinerts, 307 natural convection cooling, 308320 properties of liquids, 307 Low Reynolds number (near-wall) effects, 88-90,93 Mach number, I I Manufacturing processes, 258 casting, 260, 287 continuous casting. 260, 289 crystal growing, 258 gas cutting, 258 heat treatment, 258, 280 hot rolling, 258 metal forming, 258 optical fiber drawing, 260, 283 plastic injection molding, 258 screw extrusion, 260, 276, 283 soldering, 258 welding, 258 Materials processing, 258-295 boundary conditions for, 268-273 Dirichlet (function), 268, 270 mixed, 268 Neumann (gradient), 268, 270 Matrix bandwidth, 170 Meteorological modeling, 354 Method of characteristics, 29 Method of weakest descent, 138 Metric coefficients, 41 1 in various orthogonal coordinate systems, 41 5- 41 6 Metrics, 26 Metric tensor, 121 Mixing-plane approach, 238 Morkovin compressibility hypothesis, 91 Moving material, 274, 285, 290 Moving source, 275

42 1

index

Natural coordinate system, 54 Navier-Stokes equations, 5 parabolized (PNS), 2 1 Near-wall effects, 88-90,93 Neck-down profile, 269 Newtonian fluid, 4 Nodes, 49 exterior, 54 interior, 54 Non-Newtonian fluids, 278 Nozzle flows, 179-2 19 NPARC Navier-Stokes code, 206-207

Residual, 50, 188 Reynolds averaged equations, 7 Reynolds averaging, 76-78 Reynolds’ decomposition, 338 Reynolds number, 11 Reynolds (turbulent) flux, 78, 104 Reynolds (turbulent) stress, 78,92 RNS3D PNS code: vortex generators, 191 pressure field, 191- 192, 197 geometry specification, 192-1 93 Rotor/stator, 224

Openings, 270-27 1 Orthogonal curvilinear coordinates, 410-41 5 Orthogonal trajectory method, 130 Orthotropic materials, 28 1

Scheme: central, 62 Lax-Wendroff, 68 MacCormack, 68 Richtmyer, 68 Runge-Kutta, 68 upwind, 63 S-duct, 189-205 Secondary flow, 200,204 Sedimentation, 340 Separated flow, 182 Shape function, 49 Shear thickening (dilatant) fluids, 278-2 79 Shear thinning (pseudo-plastic) fluids, 278-279 Shock waves, 183 Solidification, 287, 293 Solution process, 181- 189 Species conservation equation, 337 Specific heat, 6 Spectral method, 37 Speed of sound, 21 Stability analysis, 44 von Neumann, 46 Stability requirement, 47 Staggered grid approach, 58 Standardization, 214-21 5 Stream function, 16 Streamline, 17 Stress tensor, 4 Stress transport models, 92-96

Panel method, 37 Parabolic equations, 16, 30 Paragon Spirit inlet, 189-205 Peclet number, turbulent, 100-101 Phase change, 264,271 Phase-change materials (PCM), 320-32 1 desirable properties of, 320 Plastic quad flat package (PQFP), 322 Point collocation method, 51 Point-in-polygon check, 169 Pollutant dynamics, 337 Pollutant species, 336 Positive-definite solution, 344-345 Postprocessing, 189,214 Power-law fluids, 279 Prandtl number, 9 turbulent, 80-81,98-101 Predictor-corrector method, 68 Printed wiring board (PWB), 322 RANS code, 224 Real-gas effects, 183 Receptor areas, 336 Reference conditions, 184 Renormalization group (RNG) models, 82, 85, 91, 94, 101

Index

422 [Stress transport models] algebraic stress models (ASM), 94-96 differential stress models (DSM), 92-94 Stretching function, 140 Strong conservative form, 25 Strouhal number, 273 Strut-jet engine, 205-21 1 Subsonic inlet, 189-205 Substantive derivative, 3 Substructuring, 56 Surface force, 3 Sutherland’s law, 8 System of equations: sparse, 43 tridiagonal, 43 System simulation, 293-294 Taylor series, 38 Tension-torsion spring analogy, 153 Thermal conductivity, 6 Thermal stratification, 3 16 Thin shear-layer equations, 12 Thomas algorithm, 43 Time-splitting approach, 341 Total pressure loss, 227 Transfinite interpolation, 145 Transformation : hinge point, 130 Schwarz-Christoffel, 128 Theodorsen-Garrick, 127 von Karman-Trefftz, 125 Transition, 223

Transition modeling, 182, 2 1 1 Transport models, 339 Turbulent diffusivity tensor, 338 Turbulent (eddy) diffusivity, 80, 102 Turbulent (eddy) viscosity, 80, 83-87, 102 Turbulent (Reynolds) flux, 78, 104 Turbulent (Reynolds) stress, 78, 92 Turbulent scales, 102 Two-boundary technique, 141 Two-phase method, 265-266 Unsteady flow, 213 Unsteady heat load, 222 User friendliness, 213,214 Validation, 2 15, 2 16 Van Driest damping, 83 Variable properties, 277, 283 Velocity potential, 19 Velocity potential equation, 20 Velocity-pressure coupling, 3 11 SIMPLER algorithm, 266,311,325 Viscoelastic fluids, 278 Viscoinelastic fluids, 278-279 Viscosity variation, 278 Viscous dissipation, 264, 279 Viscous effects, 182 Vortex generators, 189- 191,200-204 Vortices, 25 1 Vorticity, 19 Wall functions, 88 Wilcox k - o model, 87

Applied Computational Fluid Dynamics (Mechanical Engineering Series) (Dekker Mechanical Engineering) - PDF Free Download (2024)